DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing

DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing Haoxiang Cao1,2,§

Jiajiong Cao2

Xuanpu Zhang2,§

1 South China Normal University

Changqian Yu2,‡ 2 KlingAI Research

Case B (Planner Fails, Renderer Executes Wrong Plan)

arXiv:2608.20161v1 [cs.AI] 20 Aug 2026

Case A (Planner Correct, Renderer Fails)

Planner (VLM) Renderer (Diffusion)

Planner (VLM) Combine the red liquid and yellow liquid, resulting in a orange liquid filling both bowls ...

Combine the red and yellow liquids, resulting in an orange liquid filling one bowl.

Credit Assignment: Renderer > Planner

🔥 Renderer-focused Update

Chaoqun Wang1,†

Renderer (Diffusion)

Credit Assignment: Planner > Renderer Credit Assignment Weight

Credit Assignment Weight

Planner-focused Update 🔥

Cross-module Credit Assignment

Figure 1: Two visually poor editing outcomes can require opposite cross-module update emphases. Case A: the planner captures the requested edit reasonably well, but the renderer fails to realize it, so this sample should place more corrective update weight on the renderer. Case B: the sampled plan is a less useful intermediate for the instruction, and the renderer faithfully executes it, so this sample should place more corrective update weight on the planner. DARS is designed to distinguish render-dominant from plan-dominant update needs through routing signals; when planner-side emphasis is needed, its structured slots support finer within-planner diagnosis.

Abstract

1

Instruction-based image editing uses a planner-renderer pipeline: a vision-language model (VLM) first converts the instruction into an edit plan, and a diffusion model then executes that plan. Training such systems with only final-image rewards is inefficient because a poor edit does not reveal whether additional optimization should place more emphasis on the planner or the renderer, and even planner-dominant cases remain difficult to localize within a freeform reasoning trace. We present DARS, a reinforcement learning framework for dual-level credit assignment in this two-stage setting. Across modules, multi-plan multi-render rollouts estimate betweenplan and within-plan reward variability for soft module routing, while mean rewards across rollouts provide hardness estimates for an adaptive curriculum. Within the planner, a four-field structured reasoning output enables a prefix-gated reward and token-level advantage reweighting, turning outcome-level feedback into localized supervision. Experiments on five benchmarks show that DARS outperforms a Joint RL baseline with the same backbone, data, reward model, and rollout budget, with the largest gains on reasoning-intensive edits.

Many image editing systems [15, 16, 24, 31, 38] adopt a plannerrenderer pipeline: a vision-language model (VLM) [1, 25, 29] interprets the instruction and produces an edit plan, and a diffusion model [13, 18, 21, 33] renders the image conditioned on that plan. This decomposition improves controllability, but also turns joint optimization into a dual-level credit-assignment problem. Across modules, as illustrated in Figure 1, a poor image does not reveal whether the current sample would benefit more from planner-side or renderer-side corrective emphasis. Within the planner, even when planner-side updates are the more useful direction, the same outcome-level feedback still does not reveal which part of the plan should be corrected. The difficulty comes from the fact that the two modules fail in different ways. The planner is an autoregressive language model [26], so its errors are typically semantic or compositional. The renderer is a diffusion-based generator [9], so its failures appear as visual artifacts, preservation damage, or incomplete execution. A single post-render score [20, 34] collapses these qualitatively different errors into the same feedback, making credit assignment ambiguous. Existing RL [11, 16, 24, 32] strategies, therefore, either optimize one module at a time, leaving the other bottleneck untouched, or

‡ Project lead.

† Corresponding author.

§ Work done during internship in KlingAI Research.

Introduction

Preprint.

update both modules uniformly from the same reward. In either case, a low-scoring edit still does not indicate whether additional optimization should focus more on planning or on rendering. This makes credit assignment [10] under-specified at two levels. First, different samples should not contribute equally to planner and renderer updates. Second, even when the planner is the main bottleneck, outcome-level feedback still does not say which part of the plan is wrong: the requested modification, the preservation constraint, the scene-level objective, or the execution hint. To address this dual-level credit-assignment problem, we present DARS, a joint-RL framework for planner-renderer image editing. We sample multiple plans and multiple renderings per plan, then use the resulting reward statistics in two different ways: betweenplan and within-plan variability determine how much extra update weight is routed to the planner or renderer, while mean rewards across rollouts provide a hardness estimate for adaptive curriculum scheduling [2]. On the planner side, we replace an opaque freeform trace with a four-field structured reasoning output (Modify, Preserve, Overall, Tips), so the same rendered outcomes can support within-planner credit assignment through slot-level rewards and token-level advantage reweighting (Figure 2). In summary, our contributions are: • We formulate joint RL for planner-renderer image editing as a dual-level credit-assignment problem, spanning cross-module update allocation and within-planner reasoning diagnosis. • We propose a rollout variance decomposition that separates across-plan from within-plan variability for cross-module routing, together with a rollout-reward hardness estimate for adaptive curriculum scheduling. • We introduce a four-field structured reasoning output for the planner together with a prefix-gated reward and token-level advantage reweighting, enabling localized planner optimization from post-render feedback. Experiments on five benchmarks show that DARS is especially effective on reasoning-intensive edits [35, 40], and ablations indicate that cross-module routing and structured planner supervision contribute complementary gains.

2

Related Work

Instruction-based image editing. Instruction-based image editing [3, 8, 13, 14, 21, 22, 33] aims to modify a source image according to a natural-language instruction while preserving irrelevant content. Recent work increasingly augments editing with explicit reasoning or planning. RePlan [24] introduces region-aware planning for complex edits, while ReasonEdit [38] and UniReason 1.0 [30] strengthen editing with richer semantic and world-knowledge reasoning. Step1X-Edit-v1p2 [38], ThinkRL-Edit [15], and EditThinker [16] further show that multi-round reflection can improve editing fidelity, and ThinkGen [11] demonstrates the benefit of explicit intermediate reasoning in a coupled reasoning-generation pipeline. Taken together, these methods establish the value of introducing a planner-like stage before generation. Their main emphasis, however, is on designing stronger intermediate reasoning or refinement procedures. Our focus is different: we study how to optimize an already existing planner-renderer pipeline once a structured intermediate stage has been introduced.

Cao et al.

RL for two-stage planning-based editing. Reinforcement learning has recently been used to improve instruction faithfulness and multimodal alignment in image generation [17, 36, 41, 42] and editing. PromptRL [31] studies reward-based optimization for flow-based generation. In two-stage editing pipelines, existing RL methods differ mainly in which module is optimized: ThinkRL-Edit [15] focuses on the generation side by tuning the diffusion model and text encoder under reflective plans, EditThinker [16] concentrates optimization on the VLM reasoning module, and ThinkGen [11] alternates updates between the reasoning model and the generator. Once editing is organized as an explicit prompt-rewriting, planning, or reasoning stage followed by visual execution, the optimization problem becomes inherently two-stage: one module produces an intermediate textual plan, and another realizes it in the image space. In this setting, existing RL strategies largely follow three patterns: (i) stage-wise or alternating optimization, which separates planner and renderer updates; (ii) single-module optimization, which freezes one component and tunes the other; and (iii) uniform joint optimization, which updates both modules from the same outcomelevel reward. The first two reduce the opportunity for coordinated co-improvement, while the third leaves the learning signal underspecified because a low final score does not reveal which module should receive more optimization emphasis. Our method is most closely related to the third line of work, but differs in explicitly addressing this module-allocation problem through rollout-derived routing while retaining joint training. Curriculum learning and structured reasoning. Curriculum learning [4, 5] improves optimization by ordering or weighting training samples according to difficulty. Standard curriculum designs, however, treat hardness as a one-dimensional property of each sample. In two-stage image editing, this abstraction is incomplete: two samples with similar difficulty may still differ in whether their uncertainty is primarily plan-dominant or render-dominant, suggesting that sample scheduling alone is insufficient without moduleaware weighting. On the reasoning side, PromptEnhancer [32], ThinkGen [11], and EditThinker [16] show the value of richer intermediate text, but free-form reasoning traces remain difficult to evaluate locally after rendering. They often depend on engineered cold-start formats or expert-designed prompting recipes [12], which makes them harder to scale and optimize consistently under RL. Our method occupies a different point in this design space: instead of relying on long free-form traces, we use a compact four-field schema that remains close to the planner’s native generation format while exposing semantically aligned slots for post-render reward decomposition and token-level credit assignment.

3 Preliminary 3.1 Problem Setup The input is x = (𝐼 src, 𝑐), where 𝐼 src is the source image and 𝑐 is the editing instruction. A VLM planner 𝑞𝜃 (e | x) with parameters 𝜃 predicts a structured edit plan   e = 𝑒 mod, 𝑒 pre, 𝑒 ovr, 𝑒 tip , (1) specifying modifications, preservation constraints, the scene objective, and execution hints. A conditional image generator (renderer)

DARS for Instruction-Based Image Editing

(a) Free-form planner output. Draw what it will look like after the question mark is replaced with the correct answer.

(b) Structured planner produces a four-field answer.

Preprint. <think> On the blackboard, the math problem “1+1=?” is written. The question mark after the equal sign indicates that the answer is not provided, and the goal is to give the answer to this simple arithmetic problem.</think> <Imgen> Clearly write the number “2” on the blackboard. </Imgen> <Imgen>: right

<Imgen>: right

<Imgen>: right

<Imgen>: right

<Imgen>: right

<Modify>: wrong <Preserve>: <Overall>: <Tips>: -

<Modify>: wrong <Preserve>: <Overall>: <Tips>: -

<Modify>: right <Preserve>: wrong <Overall>: <Tips>: -

<Modify>: right <Preserve>: right <Overall>: right <Tips>: wrong

<Modify>: right <Preserve>: right <Overall>: right <Tips> : right

<Modify>The math problem “1+1=?” is written on the blackboard. Precisely replace the “?” with the number “2”.</Modify> <Preserve> Preserve the teacher's smile, light blue shirt, chalk-holding gesture, blackboard wooden frame, and classroom background. </Preserve> <Overall> A smiling male teacher stands in front of a chalkboard displaying the simple equation "1+1=2". </Overall> <Tips> - Retain natural chalk dust edges and uneven pressure variations in the strokes, avoiding overly smooth lines. - The spacing and baseline alignment between the number "2" and the question mark should match the original handwriting habits. - The shadow direction and the diffuse lighting on the blackboard should be consistent with the overall light source of the image. - Pay attention to the subtle reflections of chalk on the blackboard and the layering texture.</Tips>

Figure 2: Free-form vs. structured planner output. With a free-form planner trace, post-render feedback remains attached to a single opaque text sequence, making within-planner credit assignment difficult. Our four-field structured answer (Modify, Preserve, Overall, Tips) exposes semantically distinct slots, so the same rendered rollouts can be diagnosed field by field and converted into slot-wise planner credit signals. 𝑝𝜙 (y | 𝐼 src, e) with parameters 𝜙 produces the image y. The renderer receives the source image and the structured plan; the raw instruction 𝑐 is not separately passed to it. For reward computation, Gemini 3 Pro is used once before RL training to generate
and cache a slot-aligned checklist C(x) = 𝐶 mod, 𝐶 pre, 𝐶 ovr, 𝐶 tip from the source image and raw instruction. During RL training, Qwen3VL-32B performs all online planner-side and renderer-side Yes/No scoring for the sampled rollouts. Each checklist field enumerates verifiable requirements for the corresponding planner slot. The reward model produces two module-targeted scores from each rendered outcome. The renderer reward 𝑅rend (x, e, y; C) measures checklist satisfaction by the edited image. The planner reward 𝑅plan (x, e, y; C) evaluates whether each structured slot is semantically appropriate and supported by the realized edit, so plan quality is assessed in the execution context. A shared reward

𝑅share = 𝜂 plan 𝑅plan + 𝜂 rend 𝑅rend

(2)

uses fixed mixing weights 𝜂 plan, 𝜂 rend ≥ 0 for routing and curriculum estimation, while 𝑅plan and 𝑅rend remain the optimization targets. All in-house ablations use the same reward model, so their differences reflect the proposed credit-assignment and planner-structure choices under a fixed reward pipeline. The central challenge is credit assignment: for a given sample, (i) which update path should receive a stronger learning signal, and (ii) which parts of the planner output, if any, should be corrected.

3.2

Policy Optimization Background

GRPO [28] samples 𝐺 candidate outputs per input and uses grouprelative advantages with clipped surrogate and KL regularization. Flow-GRPO [17] extends this to diffusion-flow generators [6, 19] by defining the policy ratio at each denoising step. Both are used off-the-shelf; our contribution lies in the rollout-derived signals that weight their updates.

4

Method

DARS addresses credit assignment at two levels in two-stage image editing pipelines: across modules and within the planner. The framework contains three coupled components: (1) Structured reasoning and slot-wise planner credit assignment (Section 4.1): the planner outputs a four-field answer enabling a prefix-gated reward and finer diagnostics; (2) Rollout variance decomposition for cross-module credit assignment (Section 4.2): multi-plan and multi-render rollouts produce variance-based statistics that distinguish across-plan variability from within-plan rollout variability; (3) Adaptive curriculum and soft module routing (Section 4.3): a rollout-derived hardness score controls sample weighting, while decomposed variability guides how extra update weight is assigned across modules. The underlying principle is to separate three questions that end-toend editing RL conflates: which module should receive extra corrective emphasis (cross-module credit assignment via module routing), how planner errors should be diagnosed more finely (structured reasoning with slot-wise planner credit assignment), and which samples are currently most informative (adaptive curriculum). Figure 3 provides an overview.

4.1

Structured Reasoning and Slot-Wise Planner Credit Assignment

Structured reasoning output. Open-ended planner reasoning is difficult to optimize under sparse editing rewards because feedback arrives only after image generation. Following Section 3, we represent the planner
output as a four-field structured answer e = 𝑒 mod, 𝑒 pre, 𝑒 ovr, 𝑒 tip . Each field has a task-specific role: 𝑒 mod (Modify) states what should change; 𝑒 pre (Preserve) states what must remain unchanged; 𝑒 ovr (Overall) captures scene-level coherence; 𝑒 tip (Tips) provides short execution hints for the renderer. Each field remains free-form natural language, preserving the VLM backbone’s expressiveness while providing semantic anchoring for reward decomposition. For every training input, a slot-aligned

Preprint.

Cao et al.

Structured Planning & Multi-Rollout Sampling

Within-Planner Credit Assignment

Reward Evaluation & Credit Signals

“change the lamp shade from white to blue”

which samples to learn

K Renderings per Plan

Input Sample

Sample hardness

Renderer

&

Editing Instruction

Hardness ! 𝐱 𝐻

(Diffusion-Based Image Editor) Source Image

Planner (Vision-Language Model)

M Candidate Plans Plan 1: Swap the white …

Modify lamp shade color … Modify lamp shade color … Modifyroomlamp shade color … Preserve geometry, desk … Preserve room geometry, desk … lamp texture desk … Preservechange room geometry, Overall change lampposition texture … keeping Overall whilechange lampposition texture … while keeping Overallwarm glow, tidy , soft while keeping position … Tips warm glow, tidy , soft beige wall background… Tips warm glow, tidy, soft Tips beige wall background… beige wall background…

✓/✘ ✓/✘ ✓/✘

Plan 3: Take off the white …

Credit to Renderer !!"#$ 𝐱 𝑈

relatively simple

Reward

more reasoningintensive

𝑤!"# (𝐱, 𝑛) Cross-Module Credit Assignment where to learn

𝛼 𝐱

✓/✘

Slot-wise reward & advantage

Adaptive Curriculum

Credit to Planner !%&'# 𝐱 𝑈

Plan 2: Replace the white …

…

Credit Assignment & Optimization

𝜔 𝐱

Planner GRPO

🔥

🔥

Renderer Flow-GRPO

Figure 3: Framework of DARS. The structured planner produces a four-field answer; multiple plans and renderings per plan are sampled. Mean rewards across rollouts yield hardness estimates for adaptive curriculum scheduling, while reward variability is decomposed into across-plan and within-plan components to produce module-routing signals for module-specific weighting. Prefix gating further provides slot-wise planner credit signals for finer within-planner diagnosis. checklist C(x) serves as the shared evaluation anchor for both module rewards.

The renderer reward focuses on execution quality: 4 ∑︁

𝛾 𝑗𝑢 ( 𝑗 ), (4) Prefix-gated reward. With structured planner outputs, feedback 𝑗=1 can be decomposed at the slot level. For each field 𝑗 ∈ {mod, pre, ovr, tip}, the reward model inspects the slot text 𝑒 ( 𝑗 ) , the source image and where 𝑢 ( 𝑗 ) ∈ [0, 1] is analogously the unweighted fraction of instruction, the edited image y, and the corresponding checklist renderer-side Yes judgments over all questions in 𝐶 ( 𝑗 ) , and 𝛾 𝑗 ≥ 0 ( 𝑗 ) ( 𝑗 ) field 𝐶 . Encoding Yes as 1 and No as 0, the slot score 𝑠 ∈ [0, 1] is the corresponding renderer-side weight. The shared reward is is the unweighted fraction of Yes judgments over all questions in 𝑅share (x, e, y) = 𝜂 plan 𝑅plan + 𝜂 rend 𝑅rend . (5) 𝐶 ( 𝑗 ) . It measures semantic adequacy, specificity, and whether the realized edit supports that field as executable guidance. These scores Token-level advantage reweighting. For each plan e𝑖 and renderare planner-side post-render diagnostics: they assess plan quality (𝑗) ing y𝑖,𝑘 , the reward model returns per-render slot scores 𝑠𝑖,𝑘 , which through realized renderings, so a textually plausible but unexeenter the per-render prefix reward 𝑅plan via Eq. 3. The plan-level cutable plan does not receive high reward. Averaging each plan’s plan reward 𝑟𝑖 averages the resulting per-render rewards across 𝐾 score over 𝐾 independent renderings reduces renderer noise. renderings. For token-level advantage modulation, the slot scores The four fields admit a natural causal ordering: modification Í𝐾 ( 𝑗 ) intent (𝑒 mod ) is fundamental; preservation constraints (𝑒 pre ) are are likewise averaged: 𝑠¯𝑖( 𝑗 ) = 𝐾1 𝑘=1 𝑠𝑖,𝑘 . Tokens in slot 𝑗 then use meaningful only if modification is correct; the overall objective ( 𝑗 −1 plan (𝑒 ovr ) synthesizes both; tips (𝑒 tip ) are useful only when preceding Ö 𝑠¯( 𝑗 ) , 𝐴ˆ𝑖 ≥ 0, plan plan (𝑙 ) 𝐴ˆ𝑖,𝑗 = 𝐴ˆ𝑖 · 𝜎¯𝑖 · 𝑖 ( 𝑗 ) (6) fields are sound. We encode this by prefix gating. For a realized plan  (𝑗)  1 − 𝑠¯𝑖 , 𝐴ˆ𝑖 < 0, (𝑗) 𝑙=1 plan–rendering pair (e𝑖 , y𝑖,𝑘 ), let 𝜎𝑖,𝑘 = I 𝑠𝑖,𝑘 ≥ 𝛿 𝑗 , where 𝛿 𝑗 is the slot threshold and 𝜆 𝑗 ≥ 0 weights slot 𝑗. The per-render planner plan where 𝐴ˆ𝑖 is the group-relative plan-level advantage and 𝜎¯𝑖(𝑙 ) = reward is (𝑙 ) I[¯𝑠𝑖 ≥ 𝛿𝑙 ] is an analogous prefix gate computed from renderingaveraged slot scores. It uses the same threshold form and causal 𝑗 −1 4 (𝑙 ) ∑︁ Ö ordering as Eq. 3, but is distinct from the per-render gates 𝜎𝑖,𝑘 . (𝑗) (𝑙 ) 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C) = 𝜆 𝑗 · 𝑠𝑖,𝑘 · 𝜎𝑖,𝑘 , (3) This sign-aware modulation prevents low-quality slots from inher𝑗=1 𝑙=1 iting the positive credit of an otherwise good plan, while assigning stronger negative updates to the slots most associated with poor Î where slots are ordered as 𝑗 = 1:mod, 2:pre, 3:ovr, 4:tip. We use overall plan performance. The prefix gate 𝑙 < 𝑗 𝜎¯𝑖(𝑙 ) ensures that Î0 the empty-product convention 𝑙=1 (·) = 1, so the Modify term is downstream slots receive gradient only when all upstream slots always ungated. Prefix gating activates later-slot rewards only after pass their quality thresholds, maintaining the same causal depenearlier slots are correct, encouraging the planner to fix fundamental dency enforced in the reward (Eq. 3) through to the token-level mistakes first. update. 𝑅rend (x, e, y; C) =

DARS for Instruction-Based Image Editing

4.2

Preprint.

Cross-module credit assignment. We use the relative magnitude of the variance signals:

Rollout Variance Decomposition for Cross-Module Credit Assignment

For a fixed input x, let e ∼ 𝑞𝜃 (· | x) and y ∼ 𝑝𝜙 (· | 𝐼 src, e). We decompose the variance of 𝑅share :   𝑈 plan (x) = Vare Ey|𝐼src ,e [𝑅share ] , (7) h i 𝑈 rend (x) = Ee Vary|𝐼src ,e (𝑅share ) . (8) By the law of total variance, Var(𝑅share | x) = 𝑈 plan (x) + 𝑈 rend (x). The first term measures how much the expected reward changes across plans; the second measures reward fluctuation under a fixed plan. Large across-plan variability suggests that selecting or improving the plan could materially change the outcome; large within-plan variability suggests that outcome quality remains unstable even after the plan is fixed. Since the edited image varies across renderer rollouts and both 𝑅plan and 𝑅rend are derived from that realized image, within-plan variability is induced by renderer stochasticity but can propagate into both module-targeted scores. We use these quantities as rollout-level cross-module credit-assignment signals for where additional optimization is currently most actionable; accordingly, 𝑈 plan and 𝑈 rend are referred to as plan-dominant and render-dominant variability. For each sample, we draw 𝑀 plans and 𝐾 renderings per plan, share . The per-plan mean is 𝑟¯ = 1 Í 𝑟 share and yielding rewards 𝑟𝑖,𝑘 𝑖 𝑘 𝑖,𝑘 𝐾 Í the grand mean is 𝑟¯ = 𝑀1 𝑖 𝑟¯𝑖 . We compute: 𝑀

2 𝑠 within (x) =

𝐾

2 1 ∑︁ 1 ∑︁ share 𝑟𝑖,𝑘 − 𝑟¯𝑖 , 𝑀 𝑖=1 𝐾 − 1

(9)

𝛼 (x) =

bplan + 𝜀𝑈 𝑈 , brend + 2𝜀𝑈 𝑈bplan + 𝑈

𝜔 (x) = 1 − 𝛼 (x).

(13)

Here 𝜀𝑈 > 0 is the routing smoothing constant. This soft routing acts as a residual reweighting on top of the shared curriculum weight: every sample still updates both modules, while the variance signals control the extra module-specific weight. Curriculum determines when a sample receives larger weight; module routing determines where additional weight is placed. We validate these routing weights through both agreement with GPT-5 pseudo-labels (Section 5.5) and downstream ablation (Section 5.4). How curriculum and credit assignment interact. The two mechanisms play complementary roles over time. Early in training, the curriculum boundary 𝜅 (𝑛) is low, so easier samples receive most of the weight; for these samples, module routing fine-tunes which module gets emphasis. As training progresses and 𝜅 (𝑛) rises, harder samples with heterogeneous variance are admitted. Module routing then becomes more important, because these newly admitted samples do not benefit equally from additional updates to both modules. If a sample has high hardness but near-zero total variance, curriculum still controls its admission while module routing contributes little differential emphasis; optimization is then driven mainly by the two task rewards 𝑅plan and 𝑅rend .

4.4

Joint Optimization

𝑘=1

2 𝑠 between (x) =

𝑀 ∑︁

2 1 𝑟¯𝑖 − 𝑟¯ . 𝑀 − 1 𝑖=1

(10)

2 The within-plan estimate is 𝑈brend = 𝑠 within , and the across-plan estimate corrects for finite-𝐾 noise:
. (11) 𝑈bplan (x) = max 0, 𝑠 2 − 1 𝑠2 between

𝐾 within

The 𝑀 × 𝐾 rollouts are shared with policy optimization, so variance estimation requires no extra renderer budget.

4.3

Adaptive Curriculum and Soft Module Routing

Hardness determines when a sample should contribute strongly; rollout-variance signals determine where additional module emphasis should be assigned. Curriculum scheduling. The rollout-based hardness estimate is b(x) = − 1 Í𝑖,𝑘 𝑟 share (larger = harder). After normalization, the 𝐻 𝑀𝐾 𝑖,𝑘 e(x) is compared against a step-dependent resulting hardness 𝐻 boundary 𝜅 (𝑛) that increases with training step 𝑛 via the empirical e in a running buffer. The curriculum weight is quantile of 𝐻 ! e(x) 𝜅 (𝑛) − 𝐻 𝑤 cur (x, 𝑛) = sigmoid . (12) 𝜏c Here 𝜏c is a temperature parameter controlling how sharply the curriculum weight transitions around the current boundary 𝜅 (𝑛). Early in training, simpler samples receive large weights; as the model improves, 𝜅 (𝑛) rises to progressively expose harder samples.

The planner and renderer operate in different output spaces and are optimized with different policy-gradient objectives: text-GRPO [28] for the planner and flow-GRPO [17] for the renderer. The key modification relative to standard GRPO is that each module’s objective is weighted by the curriculum factor 𝑤 cur (x, 𝑛) plus a residual modulerouting emphasis. Every sample contributes a base update to both modules, scaled by its curriculum weight; module routing only adds module-specific weight according to the surrogate decomposition. For planner optimization, each plan e𝑖 receives a renderer-averaged Í plan plan reward 𝑟𝑖 = 𝐾1 𝑘 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C), from which 𝐴ˆ𝑖 is computed. Tokens in slot 𝑗 then use the sign-aware slot-weighted advantage from Eq. 6: positive plan-level advantages reinforce high-scoring slots, while negative plan-level advantages concentrate corrective updates on low-scoring slots. For renderer oprend = timization, each rendered sample keeps its own reward 𝑟𝑖,𝑘 𝑅rend (x, e𝑖 , y𝑖,𝑘 ; C). The objectives are: h i
plan Jplan (𝜃 ) = Ex 𝑤 cur 1 + 𝜌 𝛼 LGRPO (𝜃 ; x) , (14) h i
rend Jrend (𝜙) = Ex,e 𝑤 cur 1 + 𝜌 𝜔 Lflow-GRPO (𝜙; x, e) , (15) plan

where 𝜌 ≥ 0 controls the strength of module routing. LGRPO is the standard clipped surrogate with KL regularization, applied at plan rend the token level using 𝐴ˆ𝑖,𝑗 for tokens in slot 𝑗, and Lflow-GRPO rend uses per-render advantages from {𝑟𝑖,𝑘 }. The final objective is J = Jplan + Jrend . The routing weights are estimated from 𝑅share , while each module’s policy gradient uses its own reward (𝑅plan or 𝑅rend ). Module routing, therefore, modulates how much each

Preprint.

Cao et al.

Table 1: Main comparison across five benchmarks. All entries are official overall scores (higher is better; compare within columns). PICA-Bench reports “simple / detailed” prompt scores. We highlight Qwen-Image-Edit-2511-based methods separately and include a controlled Joint RL + Adaptive Curriculum baseline matching DARS in backbone, data, reward model, and rollout budget. Method

KRIS-Bench

RISE-Bench

ImgEdit-Bench

GEdit-Bench-EN

PICA-Bench

59.57 61.02 61.24 62.58 63.26 66.04 66.78 68.80

10.56 13.33 10.00 11.67 15.00 14.44 18.89 18.33

3.97 3.70 3.87 4.00 4.06 4.19 4.05 4.31

7.16 6.47 6.74 7.47 6.52 6.48 7.37 7.59

60.82/65.09 60.40/66.55 51.02/60.38 56.15/60.55 57.30/61.62 55.13/55.84 61.48/68.96 59.64/67.19

64.42 55.72 71.54 72.15 80.72

17.50 9.72 18.90 25.70 27.50

4.36 3.44 4.18 4.20 4.39

6.97 6.43 7.59 7.83 7.86

63.20/72.27 52.00/54.92 61.90/70.77 63.55/72.22 64.19/72.75

Other baselines ThinkGen [11] UniREdit [7] PromptRL [31] Step1X-Edit-v1p2 [38] UniReason 1.0 [30] ThinkRL-Edit [15] PhysicEdit [39] EditThinker [16] Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 [33] RePlan [24] PromptEnhancerV2 [32] Joint RL + Adpt. Curriculum (ctrl.) DARS

module is updated without changing what it learns. Full expanded expressions are in the supplementary.

5 Experiments 5.1 Experimental Setup Training. We train DARS on 10K examples (5K from the RL split of THINKEDIT-140K [16] and 5K from UniREdit-Data-100K [7]). The planner is initialized from Qwen3-VL-4B-Instruct, and the renderer from Qwen-Image-Edit-2511. We use 𝑀=4 plans and 𝐾=4 renderings per plan (𝑀 × 𝐾 = 16 rollouts per input). The learning rate is 2 × 10−6 , the global batch size is 256, the KL penalty coefficient [27] is 10−3 , and the maximum image resolution is 1024 × 1024. The offline checklist generator (Gemini 3 Pro) and the online reward judge (Qwen3-VL-32B) are fixed across DARS and all in-house ablations, so differences reflect the proposed credit-assignment and planner-structure choices under a fixed reward pipeline. The controlled baseline and DARS use the same rollout budget on identical hardware; DARS adds lightweight rollout statistics and slotwise scoring on top of that shared budget. We evaluate on five complementary benchmarks spanning reasoning-intensive editing, general-purpose instruction following, preservation-sensitive localized editing, and physics-aware realism: KRIS-Bench [35], RISEBench [40], ImgEdit-Bench [37], GEdit-Bench-EN [18], and PICABench [23]. All reported numbers are official overall scores, and the higher the number, the better. Benchmark results are obtained using each benchmark’s official evaluation code and are fully independent of the training reward model. Baselines. We compare against 11 methods spanning multi-round reflection editors (ThinkRL-Edit [15], Step1X-Edit-v1p2 [38], and EditThinker [16]), reasoning- or planning-based editors (ThinkGen [11], RePlan [24], UniREdit [7], and UniReason 1.0 [30], which includes

a single reflection step), reward-optimized or prompt-enhancement methods (PromptRL [31] and PromptEnhancerV2 [32]), a physicsspecialized baseline (PhysicEdit [39]), and a strong end-to-end editor (Qwen-Image-Edit-2511 [33]). We further include a controlled Joint RL + Adaptive Curriculum baseline that matches DARS in backbone, data, reward model, rollout budget, and adaptive curriculum, while replacing DARS’s structured reasoning planner and dual-level credit assignment with free-form reasoning. Detailed architecture descriptions are in the supplementary.

5.2

Main Results

Table 1 shows that the full DARS system attains the best score on all five benchmark regimes. Against previously reported methods, the gains are consistent across evaluation criteria. Against the in-house controlled baseline with matched backbone, data, reward model, rollout budget, and adaptive curriculum, DARS improves by +8.57 on KRIS-Bench, +1.80 on RISE-Bench, +0.19 on ImgEdit-Bench, +0.03 on GEdit-Bench-EN, and +0.64/+0.53 on PICA-Bench. The strongest gains appear on KRIS-Bench and RISE-Bench, consistent with DARS’s focus on planner-renderer credit assignment for reasoning-intensive edits.

5.3

Qualitative Results

Figure 4 compares DARS against reasoning-oriented editing methods. We include Qwen-Image-Edit-2511 as a baseline without explicit reasoning, together with instruction-rewriting or planning baselines: UniReason 1.0 based on the unified multimodal model Bagel, PromptRL as a VLM-diffusion joint-RL method, EditThinker with up to five rounds of reflective prompt revision, RePlan with region-aware reasoning, PhysicEdit specialized by physics-grounded data tuning, and PromptEnhancerV2, a strong 32B prompt-enhancement model combining expert-designed rules with end-to-end RL. Across

DARS for Instruction-Based Image Editing

Instruction

Draw what it will look like after 30 seconds in summer.

Preprint.

Draw what it will look like ten seconds later.

The caterpillar wants Draw what it to find the shortest Remove the water path to the leaf. will look like Please mark the two from the after 30 shortest paths in the glass cup. minutes. diagram.

Move only two matchsticks on the left side of the equation to make the equation true. Draw the final equation.

Ori image

Qwen-ImageEdit-2511

UniReason

PromptRL

EditThinker

RePlan

PhysicEdit

Prompt EnhancerV2

Ours

Figure 4: Qualitative comparison on reasoning-intensive editing tasks requiring temporal, spatial, logical, physical, and world-knowledge reasoning, often in combination. Compared with baseline methods, DARS achieves the strongest overall performance, producing more faithful edits with better preservation and stronger scene-level consistency. the Figure 4 examples, DARS delivers the strongest visual quality, achieving the best balance of instruction faithfulness, preservation, scene coherence, and reasoning consistency.

5.4

Ablation Studies

Unless otherwise stated, ablations are reported on RISE-Bench and GEdit-Bench-EN. All variants use DARS without auxiliary heuristics. Free-form joint RL without a curriculum diverged, whereas the structured reasoning planner remained trainable without it, albeit

with worse performance. The ablations below, therefore, use the structured reasoning planner unless explicitly labeled otherwise.

Curriculum and routing (Table 2). Adaptive curriculum consistently outperforms static and no-curriculum baselines, confirming that dynamic hardness re-estimation is critical as model competence evolves. Notably, No Curriculum still converges under the structured reasoning planner, whereas free-form joint RL without a curriculum diverges , suggesting that the four-field structure provides implicit training stabilization. For routing, soft continuous

Preprint.

Cao et al.

Table 2: Ablation on curriculum and routing strategies. Variant

RISE

GEdit Variant

RISE

GEdit

7.12 7.49 7.86

Planner structure Free-form Reasoning Modify Only Without Preserve Without Overall Without Tips Full Structured Answer

23.94 25.21 26.08 26.42 26.87 27.50

6.98 7.11 7.34 7.51 7.69 7.86

7.10 7.83 7.86

Reward composition Flat Average Weighted Sum Prefix-Gated (Ours)

24.92 26.31 27.50

7.28 7.57 7.86

Curriculum No Curriculum Static Curriculum Adaptive Curriculum

18.89 24.56 27.50

6.87 7.15 7.86

Routing No Routing Hard Routing Soft Routing (Ours)

20.05 24.86 27.50

Configuration Structured Prompt Only (No RL) Free-form Joint RL + Adpt. Full DARS

23.80 25.70 27.50

weighting outperforms both uniform and hard-threshold alternatives, consistent with mixed but asymmetric variance patterns in many editing failures. The configuration comparison shows that the full DARS improves over both the prompt-only baseline (23.80, no RL) and the controlled free-form Joint RL baseline (25.70), confirming gains over structured prompting alone and over a matched free-form joint-RL setup. Planner structure and reward (Table 3). A monotonic trend runs from free-form reasoning to the full four-field answer, with each field providing gains: removing Preserve causes a drop on GEditBench-EN, consistent with its emphasis on localized editing and non-target preservation; removing Overall produces a larger drop on the reasoning-heavy RISE-Bench, indicating that scene-level coherence is important for complex edits; removing Tips yields the smallest consistent degradation, suggesting that execution guidance is helpful once higher-level semantics are in place. These results support distinct slot roles and slot-wise advantage reweighting unavailable under free-form reasoning. For reward composition, the prefix-gated reward outperforms flat averaging and weighted summation. Flat averaging treats all fields as independent and equally important, failing to reflect the causal structure of editing plans. Weighted summation improves over flat averaging but underperforms because it does not enforce dependency between upstream and downstream fields. The prefix-gated reward yields the best results by ensuring that later fields are only rewarded once earlier fields are reliable, reducing the incentive to compensate for upstream errors by overproducing downstream text.

5.5

Table 3: Ablation on planner structure and reward composition.

Validating the Uncertainty Signals

The modeling claim of DARS is that rollout statistics provide informative routing signals for cross-module optimization. We validate whether the plan-dominant and render-dominant routing scores align with GPT-5 pseudo-labels of dominant failure patterns. Protocol. For each evaluation case, we construct the same 𝑀 × 𝐾 rollout bank used by DARS and compute 𝑈bplan , 𝑈brend , and 𝛼 from that rollout bank. Separately, we sample five realized triplets from the same case and ask GPT-5 to assign each one a pseudolabel among planner-dominant, renderer-dominant, or mixed; the

Table 4: Routing-signal discrimination against majorityvoted GPT-5 pseudo-labels. Setting Planner vs. Renderer Planner vs. Mixed Renderer vs. Mixed

Accuracy

Macro-F1

AUROC

86.0 84.7 85.3

85.8 84.1 84.9

0.930 0.918 0.923

majority vote becomes the case-level pseudo-label. GPT-5 sees only the source image, structured plan, and edited image for each triplet, not the rollout statistics or reward-model outputs used by DARS. Metrics are reported for pairwise one-vs-one discrimination tasks, with AUROC computed in the corresponding binary setting. Table 4 shows AUROC above 0.91 in all pairwise discriminations, indicating that the rollout statistics track GPT-5 pseudo-labels of dominant failure patterns well. Combined with the ablation result that soft routing outperforms both hard routing and no routing (Table 2), this provides evidence that the routing signal captures failure-mode differences and improves update allocation.

6

Limitations

DARS requires 𝑀 ×𝐾 rollouts per training example, so the total cost remains higher than single-path updates even though these rollouts are shared with policy optimization. The variance-based routing and post-render planner scores depend on the reward model, and planner-renderer interactions can be harder to disentangle when both modules fail simultaneously. Adapting the framework to video editing or multi-turn settings is left for future work.

7

Conclusion

We present DARS, which formulates joint RL in two-stage plannerrenderer image editing pipelines as a dual-level credit-assignment problem. Rollout variance decomposition provides routing signals for cross-module updates; hardness-derived scores drive adaptive curriculum scheduling; and a structured four-field planner with prefix gating enables slot-wise diagnostics. DARS attains the best results across five benchmarks, with the largest gains on reasoningintensive tasks.

DARS for Instruction-Based Image Editing

Acknowledgments This work was supported by KlingAI Research. We thank Shenhui Zhang, Jiahao Guo, and Bao Tang for their valuable discussions.

References [1] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-VL Technical Report. CoRR abs/2502.13923 (2025). [2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (Montreal, Quebec, Canada) (ICML ’09). Association for Computing Machinery, New York, NY, USA, 41–48. [3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. 2023. InstructPix2Pix: Learning to Follow Image Editing Instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. IEEE, 18392–18402. [4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, Nicu Sebe, and Mubarak Shah. 2025. Curriculum Direct Preference Optimization for Diffusion and Consistency Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025. Computer Vision Foundation / IEEE, 2824–2834. [5] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, Nicu Sebe, and Mubarak Shah. 2026. Curriculum-DPO++: Direct Preference Optimization via Data and Model Curricula for Text-to-Image Generation. CoRR abs/2602.13055 (2026). [6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. 2024. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 (Proceedings of Machine Learning Research). PMLR / OpenReview.net, 12606– 12633. [7] Feng Han, Yibin Wang, Chenglin Li, Zheming Liang, Dianyi Wang, Yang Jiao, Zhipeng Wei, Chao Gong, Cheng Jin, Jingjing Chen, and Jiaqi Wang. 2025. UniREditBench: A Unified Reasoning-based Image Editing Benchmark. CoRR abs/2511.01295 (2025). [8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. 2023. Prompt-to-Prompt Image Editing with Cross-Attention Control. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. [9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. [10] Guochao Jiang, Wenfeng Feng, Guofeng Quan, Chuzhan Hao, Yuewei Zhang, Guohua Liu, and Hao Wang. 2025. VCRL: Variance-based Curriculum Reinforcement Learning for Large Language Models. CoRR abs/2509.19803 (2025). [11] Siyu Jiao, Yiheng Lin, Yujie Zhong, Qi She, Wei Zhou, Xiaohan Lan, Zilong Huang, Fei Yu, Yingchen Yu, Yunqing Zhao, Yao Zhao, and Yunchao Wei. 2025. ThinkGen: Generalized Thinking for Visual Generation. CoRR abs/2512.23568 (2025). [12] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2023. Decomposed Prompting: A Modular Approach for Solving Complex Tasks. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. [13] Black Forest Labs. 2025. FLUX.2: Frontier Visual Intelligence. https://bfl.ai/blog/ flux-2. [14] Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. 2025. FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space. CoRR abs/2506.15742 (2025). [15] Hengjia Li, Liming Jiang, Qing Yan, Yizhi Song, Hao Kang, Zichuan Liu, Xin Lu, Boxi Wu, and Deng Cai. 2026. ThinkRL-Edit: Thinking in Reinforcement Learning for Reasoning-Centric Image Editing. CoRR abs/2601.03467 (2026). [16] Hongyu Li, Manyuan Zhang, Dian Zheng, Ziyu Guo, Yimeng Jia, Kaituo Feng, Hao Yu, Yexin Liu, Yan Feng, Peng Pei, Xunliang Cai, Linjiang Huang, Hongsheng Li, and Si Liu. 2025. EditThinker: Unlocking Iterative Reasoning for Any Image Editor. CoRR abs/2512.05965 (2025). [17] Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. 2025. Flow-GRPO: Training Flow Matching Models via Online RL. CoRR abs/2505.05470 (2025).

Preprint.

[18] Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, Guopeng Li, Yuang Peng, Quan Sun, Jingwei Wu, Yan Cai, Zheng Ge, Ranchen Ming, Lei Xia, Xianfang Zeng, Yibo Zhu, Binxing Jiao, Xiangyu Zhang, Gang Yu, and Daxin Jiang. 2025. Step1X-Edit: A Practical Framework for General Image Editing. CoRR abs/2504.17761 (2025). [19] Xingchao Liu, Chengyue Gong, and Qiang Liu. 2023. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. [20] Xin Luo, Jiahao Wang, Chenyuan Wu, Shitao Xiao, Xiyan Jiang, Defu Lian, Jiajun Zhang, Dong Li, and Zheng Liu. 2025. EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling. CoRR abs/2509.23909 (2025). [21] Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, Xunliang Cai, Yayong Guan, and Jie Hu. 2025. LongCat-Image Technical Report. CoRR abs/2512.07584 (2025). [22] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. 2022. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. [23] Yuandong Pu, Le Zhuo, Songhao Han, Jinbo Xing, Kaiwen Zhu, Shuo Cao, Bin Fu, Si Liu, Hongsheng Li, Yu Qiao, Wenlong Zhang, Xi Chen, and Yihao Liu. 2025. PICABench: How Far Are We from Physically Realistic Image Editing? CoRR abs/2510.17681 (2025). [24] Tianyuan Qu, Lei Ke, Xiaohang Zhan, Longxiang Tang, Yuqi Liu, Bohao Peng, Bei Yu, Dong Yu, and Jiaya Jia. 2025. RePlan: Reasoning-guided Region Planning for Complex Instruction-based Image Editing. CoRR abs/2512.16864 (2025). [25] Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. https://qwen. ai/blog?id=qwen3.5 [26] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training. (2018). [27] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal Policy Optimization Algorithms. CoRR abs/1707.06347 (2017). [28] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. CoRR abs/2402.03300 (2024). [29] Qwen Team. 2025. Qwen3-VL Technical Report. CoRR abs/2511.21631 (2025). [30] Dianyi Wang, Chaofan Ma, Feng Han, Size Wu, Wei Song, Yibin Wang, Zhixiong Zhang, Tianhang Wang, Siyuan Wang, Zhongyu Wei, and Jiaqi Wang. 2026. UniReason 1.0: A Unified Reasoning Framework for World Knowledge Aligned Image Generation and Editing. CoRR abs/2602.02437 (2026). [31] Fu-Yun Wang, Han Zhang, Michaël Gharbi, Hongsheng Li, and Taesung Park. 2026. PromptRL: Prompt Matters in RL for Flow-Based Image Generation. CoRR abs/2602.01382 (2026). [32] Linqing Wang, Ximing Xing, Yiji Cheng, Zhiyuan Zhao, Donghao Li, Tiankai Hang, Jiale Tao, Qixun Wang, Ruihuang Li, Comi Chen, Xin Li, Mingrui Wu, Xinchi Deng, Shuyang Gu, Chunyu Wang, and Qinglin Lu. 2025. PromptEnhancer: A Simple Approach to Enhance Text-to-Image Models via Chain-of-Thought Prompt Rewriting. CoRR abs/2509.04545 (2025). [33] Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Shengming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun Wen, Wensen Feng, Xiaoxiao Xu, Yi Wang, Yichang Zhang, Yongqiang Zhu, Yujia Wu, Yuxuan Cai, and Zenan Liu. 2025. Qwen-Image Technical Report. CoRR abs/2508.02324 (2025). [34] Keming Wu, Sicong Jiang, Max Ku, Ping Nie, Minghao Liu, and Wenhu Chen. 2025. EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing. CoRR abs/2509.26346 (2025). [35] Yongliang Wu, Zonghui Li, Xinting Hu, Xinyu Ye, Xianfang Zeng, Gang Yu, Wenbo Zhu, Bernt Schiele, Ming-Hsuan Yang, and Xu Yang. 2025. KRIS-Bench: Benchmarking Next-Level Intelligent Image Editing Models. CoRR abs/2505.16707 (2025). [36] Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, and Ping Luo. 2025. DanceGRPO: Unleashing GRPO on Visual Generation. CoRR abs/2505.07818 (2025). [37] Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. 2025. ImgEdit: A Unified Image Editing Dataset and Benchmark. CoRR abs/2505.20275 (2025). [38] Fukun Yin, Shiyu Liu, Yucheng Han, Zhibo Wang, Peng Xing, Rui Wang, Wei Cheng, Yingming Wang, Aojie Li, Zixin Yin, Pengtao Chen, Xiangyu Zhang, Daxin Jiang, Xianfang Zeng, and Gang Yu. 2025. ReasonEdit: Towards ReasoningEnhanced Image Editing Models. CoRR abs/2511.22625 (2025). [39] Liangbing Zhao, Le Zhuo, Sayak Paul, Hongsheng Li, and Mohamed Elhoseiny. 2026. From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors. CoRR abs/2602.21778 (2026).

Preprint.

[40] Xiangyu Zhao, Peiyuan Zhang, Kexian Tang, Hao Li, Zicheng Zhang, Guangtao Zhai, Junchi Yan, Hua Yang, Xue Yang, and Haodong Duan. 2025. Envisioning Beyond the Pixels: Benchmarking Reasoning-Informed Visual Editing. CoRR abs/2504.02826 (2025). [41] Kaiwen Zheng, Huayu Chen, Haotian Ye, Haoxiang Wang, Qinsheng Zhang, Kai Jiang, Hang Su, Stefano Ermon, Jun Zhu, and Ming-Yu Liu. 2025. DiffusionNFT: Online Diffusion Reinforcement with Forward Process. CoRR abs/2509.16117 (2025). [42] Huaisheng Zhu, Teng Xiao, and Vasant G. Honavar. 2025. DSPO: Direct Score Preference Optimization for Diffusion Model Alignment. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net.

Cao et al.

Supplementary Material for DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing 0. Contents and Index

S7

Planner-side prompt template used during rollout to rewrite a raw editing instruction into a four-slot plan.

S8

Prompt template used in the offline Gemini 3 Pro preprocessing stage to generate the shared slot-aligned checklist C(x) from the source image and user instruction. 18

S9

Planner-side single-question prompt for the Modify slot. 19

Contents 0. Contents and Index 1 Contents 1 List of Figures 1 List of Tables 1 1 Additional Methodological Details 2 1.1 Expanded Optimization Objectives 2 1.2 Variance-Based Cross-Module Credit Assignment 2 1.3 Adaptive Curriculum and Soft Module Routing 3 1.4 Overall Training Procedure 3 2 Structured Planning and Reward Design 3 2.1 Structured Planner Output Format 3 Shared Checklist Construction and Slot 2.2 Alignment 3 2.3 Reward Instantiation in Practice 4 2.4 Illustrative Prefix-Gated Reward Example 5 3 Experimental Setup Details 7 3.1 Training Setup and Hyperparameters 7 3.2 Controlled Baseline and Evaluation Protocol 7 3.3 KRIS-Bench Breakdown 7 3.4 RISE-Bench Breakdown 7 3.5 ImgEdit-Bench Breakdown 7 3.6 GEdit-Bench-EN Breakdown 7 3.7 PICA-Bench Simple-Prompt Breakdown 7 3.8 PICA-Bench Explicit-Prompt Breakdown 7 3.9 Baseline Architecture Overview 7 4 Failure Attribution Validation Details 7 5 Additional Qualitative Results 8 5.1 Structured Plans for Qualitative Cases 12 6 Prompt Templates 12 7 Limitations 12

List of Figures S1

Five-panel schematic of checklist-based prefix-gated planner reward on a sequence-completion edit.

6

S2

Additional baseline comparisons on challenging edits.

10

S3

Representative failure cases illustrating current limitations.

10

S4

Additional reasoning-intensive editing examples.

14

S5

Additional general editing examples.

15

S6

System prompt used for GPT-5 failure-attribution annotation in the routing-signal validation study.

16

17

S10 Planner-side single-question prompt for the Preserve slot. 20 S11 Planner-side single-question prompt for the Overall slot. 21 S12 Planner-side single-question prompt for the Tips slot.

22

S13 Renderer-side single-question prompt for the Modify slot. 23 S14 Renderer-side single-question prompt for the Preserve slot.

24

S15 Renderer-side single-question prompt for the Overall slot. 25 S16 Renderer-side single-question prompt for the Tips slot.

26

List of Tables S1

Training recipe summary.

7

S2

Fine-grained KRIS-Bench results.

8

S3

Fine-grained RISE-Bench results.

8

S4

Fine-grained ImgEdit-Bench results.

9

S5

Fine-grained GEdit-Bench-EN results.

9

S6

Fine-grained PICA-Bench results under the simple-prompt setting. 11

S7

Fine-grained PICA-Bench results under the explicitprompt setting.

11

S8

Detailed baseline architecture summary.

12

S9

Actual four-slot plan decompositions for representative qualitative cases. 13

Preprint.

Cao et al.

1 Additional Methodological Details 1.1 Expanded Optimization Objectives This subsection expands the compact objectives in the main paper into the exact rollout-level forms used during optimization. Definitions of the reward signals and the high-level design motivation remain in the main text; here we focus only on how the offline cached checklist, online planner-generated four-slot plans, old-policy rollouts, and token- or step-level surrogates are instantiated in training. Throughout this subsection, 𝑖 indexes sampled plans, 𝑘 renderings, 𝑗 structured slots, 𝑡 planner tokens, and 𝑚 denoising-trajectory positions. Planner objective. For each input x, we sample a group of 𝐺 p = 𝑀 candidate plans from the behavior policy: e𝑖 ∼ 𝑞𝜃 old (· | x),

𝑖 = 1, . . . , 𝐺 p .

(S1)

For each sampled plan e𝑖 , we generate 𝐾 paired renderings y𝑖,𝑘 ∼ 𝑝𝜙 old (· | 𝐼 src, e𝑖 ). The planner reward for the 𝑖-th sampled plan is the planner reward averaged over its rendered outcomes: 𝐾

plan

𝑟𝑖

=

1 ∑︁ 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C), 𝐾

(S2)

𝑘=1

where the checklist C = C(x) is pre-generated once offline before RL training from the source image and instruction, and then shared across all rollouts of the same sample. The group-relative advantage is   plan plan 𝐺 𝑟𝑖 − mean {𝑟𝑖 ′ }𝑖 ′ p=1 plan   𝐴ˆ𝑖 = . (S3) plan 𝐺 std {𝑟𝑖 ′ }𝑖 ′ p=1 + 𝜀𝐴 Here 𝜀𝐴 > 0 is the advantage-normalization stabilizer shared by (𝑗) the planner and renderer objectives. Let 𝑠𝑖,𝑘 denote the score of slot 𝑗 under rendering y𝑖,𝑘 , and define the rendering-averaged slot score and reweighted token-level advantage as ( 𝑗 −1 𝐾 plan (𝑗) ∑︁ Ö 𝐴ˆ𝑖 ≥ 0, plan plan (𝑗) (𝑗) (𝑙 ) 𝑠¯𝑖 , 1 ˆ ˆ 𝑠¯𝑖 = 𝐾 𝑠𝑖,𝑘 , 𝐴𝑖,𝑗 = 𝐴𝑖 · 𝜎¯𝑖 · plan (𝑗) 1 − 𝑠¯𝑖 , 𝐴ˆ𝑖 < 0, 𝑘=1 𝑙=1 (S4) where 𝜎¯𝑖(𝑙 ) = I[¯𝑠𝑖(𝑙 ) ≥ 𝛿𝑙 ] is an analogous prefix gate computed from rendering-averaged slot scores. It has the same threshold form and causal ordering as the per-render gate in the main paper, but (𝑙 ) is distinct from 𝜎𝑖,𝑘 . This sign-aware modulation prevents lowquality slots from inheriting positive credit from an otherwise good plan, while assigning stronger negative updates to the slots most associated with poor overall plan performance. Since the planner is autoregressive, the GRPO ratio is defined at the token level. For the 𝑡-th token in e𝑖 : plan

𝜚 𝑖,𝑡 (𝜃 ) =

𝑞𝜃 (𝑒𝑖,𝑡 | x, 𝑒𝑖,<𝑡 ) . 𝑞𝜃 old (𝑒𝑖,𝑡 | x, 𝑒𝑖,<𝑡 )

(S5)

Let slot(𝑡) denote the structured slot containing token 𝑡, as determined by the enclosing ordered tag pair; the opening and closing tag tokens are assigned to that same slot. The token-level surrogate plan uses the sign-aware slot-weighted advantage 𝐴ˆ𝑖,slot(𝑡 ) instead of a

uniform plan-level advantage. The full planner objective is: " 𝐺p |e𝑖 |
1 ∑︁ 1 ∑︁ Jplan (𝜃 ) = Ex 𝑤 cur (x, 𝑛) 1 + 𝜌 𝛼 (x) 𝐺 p 𝑖=1 |e𝑖 | 𝑡 =1  plan plan min 𝜚 𝑖,𝑡 (𝜃 ) 𝐴ˆ𝑖,slot(𝑡 ) ,
plan
plan clip 𝜚 𝑖,𝑡 (𝜃 ), 1 − 𝜖p, 1 + 𝜖p 𝐴ˆ𝑖,slot(𝑡 ) #  − 𝛽 p 𝐷 KL (𝑞𝜃 ∥ 𝑞 ref ) .

(S6)

𝑀 from Renderer objective. Using the same sampled plans {e𝑖 }𝑖=1 the planner rollout bank, we sample 𝐾 renderer rollouts for each fixed pair (x, e𝑖 ) from the old diffusion-flow policy. Let y𝑖,𝑘 denote the decoded image from the 𝑘-th rollout under plan e𝑖 . The rendererside reward is rend 𝑟𝑖,𝑘 = 𝑅rend (x, e𝑖 , y𝑖,𝑘 ; C), (S7)

where again C = C(x) is fixed for all rollouts of the same input. The renderer-side group-relative advantage is computed separately within each fixed plan:   rend − mean {𝑟 rend }𝐾 𝑟𝑖,𝑘 ′ ′ 𝑘 =1 𝑖,𝑘 rend   𝐴ˆ𝑖,𝑘 = . (S8) rend 𝐾 std {𝑟𝑖,𝑘 ′ }𝑘 ′ =1 + 𝜀𝐴 For flow-GRPO, let 𝑚 index the complete ordered denoising trajectory, 𝜏𝑚 its diffusion time, and 𝑧𝑖,𝑘,𝑚 the corresponding latent state. Here 𝜏𝑚−1 is the adjacent lower-noise time after 𝜏𝑚 , with latent state 𝑧𝑖,𝑘,𝑚−1 ; the optimized indices form a subset S of this trajectory. The ratio is flow 𝜚 𝑖,𝑘,𝑚 (𝜙) =

𝑝𝜙 (𝑧𝑖,𝑘,𝑚−1 | 𝐼 src, e𝑖 , 𝑧𝑖,𝑘,𝑚 , 𝜏𝑚 ) . 𝑝𝜙 old (𝑧𝑖,𝑘,𝑚−1 | 𝐼 src, e𝑖 , 𝑧𝑖,𝑘,𝑚 , 𝜏𝑚 )

The full renderer objective is: " Jrend (𝜙) = Ex 𝑤 cur (x, 𝑛) 1 + 𝜌 𝜔 (x)

𝑀 𝐾
1 ∑︁ 1 ∑︁ 1 ∑︁ 𝑀 𝑖=1 𝐾 |S| 𝑘=1



(S9)

𝑚∈ S

flow rend flow rend
min 𝜚 𝑖,𝑘,𝑚 (𝜙) 𝐴ˆ𝑖,𝑘 , clip(𝜚 𝑖,𝑘,𝑚 (𝜙), 1 − 𝜖r, 1 + 𝜖r ) 𝐴ˆ𝑖,𝑘

#  − 𝛽 r 𝐷 KL (𝑝𝜙 ∥ 𝑝 ref ) . (S10)

1.2

Variance-Based Cross-Module Credit Assignment

The main paper introduces the plan-dominant and render-dominant variability terms. Here, we record only the estimator used in training. For a fixed input x = (𝐼 src, 𝑐), let e ∼ 𝑞𝜃 (· | x) and y ∼ 𝑝𝜙 (· | 𝐼 src, e). We decompose the shared-reward variance as   𝑈 plan (x) = Vare Ey|𝐼src ,e [𝑅share ] , (S11) h i 𝑈 rend (x) = Ee Vary|𝐼src ,e (𝑅share ) . (S12) By the law of total variance, Var(𝑅share | x) = 𝑈 plan (x) + 𝑈 rend (x).

(S13)

DARS for Instruction-Based Image Editing

Preprint.

For each training sample, the same 𝑀 × 𝐾 rollout bank already used for reward computation is reused to estimate these two terms. share be the shared reward of the 𝑘-th rendering under the Let 𝑟𝑖,𝑘 Í𝐾 share 𝑖-th sampled plan, 𝑟¯𝑖 = 𝐾1 𝑘=1 𝑟𝑖,𝑘 the per-plan mean, and 𝑟¯ = 1 Í𝑀 ¯ 𝑟 the grand mean. We compute 𝑖 𝑖=1 𝑀 𝑀

2 𝑠 within (x) =

𝐾

2 1 ∑︁ 1 ∑︁ share 𝑟𝑖,𝑘 − 𝑟¯𝑖 , 𝑀 𝑖=1 𝐾 − 1

(S14)

𝑀
2 1 ∑︁ 𝑟¯𝑖 − 𝑟¯ . 𝑀 − 1 𝑖=1

(S15)

𝑘=1

2 𝑠 between (x) =

The within-plan Monte Carlo estimate is 2 𝑈brend (x) = 𝑠 within (x),

(S16)

and the across-plan estimate subtracts the renderer-noise term induced by averaging only 𝐾 renderings per plan:
1 2 bplan (x) = max 0, 𝑠 2 𝑈 (S17) between (x) − 𝐾 𝑠 within (x) . The max(0, ·) truncation avoids negative across-plan estimates caused by finite-sample noise. No additional renderer calls are introduced: the same rollout bank feeds reward estimation, curriculum weighting, routing, and policy optimization.

1.4

Overall Training Procedure

Algorithm 1 summarizes one training iteration of DARS. A single shared 𝑀 ×𝐾 rollout bank is reused throughout the iteration: it first produces module-specific rewards, then yields planner-side and renderer-side advantages, and finally supplies the variance, routing, and curriculum statistics used to weight the two objectives. This reuse is important operationally because it keeps checklist loading, reward estimation, and module-specific optimization aligned to the same sampled evidence without introducing extra renderer calls beyond the rollout budget already used for policy optimization.

2 Structured Planning and Reward Design 2.1 Structured Planner Output Format The main paper already motivates the four-slot planner decomposition and the associated prefix-gated dependency structure. Here we record the operational slot semantics used throughout the supplementary examples and prompt templates: Modify specifies the requested visible change, Preserve specifies the important content that must remain unchanged, Overall specifies scene-level coherence requirements, and Tips provides localized renderer-facing execution details. The online planner serializes these free-form fields in the fixed order

Adaptive Curriculum and Soft Module Routing

<Modify>...</Modify> <Preserve>...</Preserve> <Overall>...</Overall> <Tips>...</Tips>.

This subsection records the rollout statistics used to instantiate the main-paper curriculum and routing definitions as per-sample training weights.

The matching tag boundaries determine the token-to-slot mapping slot(𝑡), with each tag pair assigned to its enclosed slot. This is the same slot schema used by the online planner output, the offline checklist cache, the reward prompts, and the qualitative plan table.

1.3

Curriculum scheduling. The rollout-based hardness estimate is 𝑀 ∑︁ 𝐾 ∑︁ share b(x) = − 1 𝐻 𝑟𝑖,𝑘 , 𝑀𝐾 𝑖=1

2.2 (S18)

Shared Checklist Construction and Slot Alignment

𝑘=1

where larger values indicate harder samples. In implementation, b(x) is normalized to 𝐻 e(x) using the running hardness statistics 𝐻 maintained over recent training batches, and the curriculum boundary 𝜅 (𝑛) is taken as an empirical quantile of that same buffer at step 𝑛. The curriculum weight is ! e(x) 𝜅 (𝑛) − 𝐻 . (S19) 𝑤 cur (x, 𝑛) = sigmoid 𝜏c Here 𝜏c controls how sharply the weight changes around the current boundary. This makes the curriculum update depend only on rollout hardness, not on which module is currently more variable. Soft module routing. We use the relative magnitude of the variance signals to define 𝛼 (x) =

bplan (x) + 𝜀𝑈 𝑈 , 𝑈bplan (x) + 𝑈brend (x) + 2𝜀𝑈

𝜔 (x) = 1 − 𝛼 (x),

(S20)

where 𝜀𝑈 > 0 is the routing smoothing constant. Routing is applied only as a residual multiplier on top of the shared curriculum factor, so every admitted sample still updates both modules. The variance decomposition only decides where the extra module-specific emphasis should go; it is not used as a hard assignment rule.

The reward pipeline uses a single slot-aligned checklist shared by planner reward, renderer reward, hardness estimation, and routing. The implementation has three distinct stages. First, during rollout, the current planner rewrites the raw instruction into a four-slot plan e; Fig. S7 shows this planner-side prompt. Second, before RL starts, Gemini 3 Pro preprocesses the training set once and directly generates a cached checklist C(x) from the source image and raw user instruction; Fig. S8 shows this offline checklist-generation prompt. Third, during RL training, Qwen3-VL-32B performs online Yes/No scoring by pairing the current rollout plan, the cached checklist, and the realized edited image. Gemini 3 Pro is not called at rollout time. Operationally, Gemini 3 Pro directly produces the shared check
list C(x) = 𝐶 mod, 𝐶 pre, 𝐶 ovr, 𝐶 tip from the source image and raw instruction, with a question list for each slot role. This preprocessing is run once before RL training, so the same cached C(x) is reused across all 𝑀 ×𝐾 rollouts of a sample. At rollout time, the current planner separately produces its own four-slot plan in the same slot schema. The checklist, therefore, defines the shared evaluation target, while the rollout-time plan provides the instance-specific slot text being judged. Each checklist field is not copied from the planner output. Instead, Gemini 3 Pro maps the source image and raw instruction

Preprint.

Cao et al.

Algorithm 1 DARS Training Algorithm Require: Planner policy 𝑞𝜃 , renderer policy 𝑝𝜙 , reference policies 𝑞 ref , 𝑝 ref , objective hyperparameters 𝜖p , 𝜖r , 𝜀𝐴 , 𝜀𝑈 , 𝛽 p , 𝛽 r , 𝜌, dataset D, rollout counts 𝑀, 𝐾, training step 𝑛 Ensure: Updated planner parameters 𝜃 and renderer parameters 𝜙 1: Refresh old policies: 𝑞𝜃 old ← 𝑞𝜃 , 𝑝𝜙 old ← 𝑝𝜙 2: Sample minibatch B ∼ D 3: for each input x ∈ B do 4: Load pre-generated shared checklist C(x) 5: for 𝑖 = 1 to 𝑀 do 6: Sample plan e𝑖 ∼ 𝑞𝜃 old (· | x) 7: for 𝑘 = 1 to 𝐾 do 8: Sample rendering y𝑖,𝑘 ∼ 𝑝𝜙 old (· | 𝐼 src , e𝑖 ) 9: Evaluate 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C), 𝑅rend (x, e𝑖 , y𝑖,𝑘 ; C), and 𝑅share (x, e𝑖 , y𝑖,𝑘 ) 10: end for 11: end for plan 𝑀 plan 𝑀 12: Compute planner-side rewards {𝑟𝑖 }𝑖=1 and group-relative advantages {𝐴ˆ𝑖 }𝑖=1 (𝑗) (𝑗) (𝑗) (𝑗) 1 Í𝐾 13: Compute rendering-averaged slot scores 𝑠¯𝑖 = 𝐾 𝑘=1 𝑠𝑖,𝑘 and prefix gates 𝜎¯𝑖 = I[𝑠¯𝑖 ≥ 𝛿 𝑗 ] for all sampled plans and slots plan plan ( 𝑗 ) ( 𝑗 ) 14: Compute sign-aware slot-weighted token advantages 𝐴ˆ from 𝐴ˆ , 𝑠¯ , and 𝜎¯ 𝑖,𝑗

𝑖

𝑖

𝑖

for 𝑖 = 1 to 𝑀 do rend } 𝐾 and group-relative advantages { 𝐴ˆ rend } 𝐾 over the 𝐾 renderings under the fixed plan e Compute renderer-side rewards {𝑟𝑖,𝑘 𝑖 𝑘=1 𝑖,𝑘 𝑘=1 17: end for bplan (x), 𝑈 brend (x), and raw hardness 𝐻 b (x) from the same 𝑀 × 𝐾 rollout bank 18: Estimate 𝑈 b (x) to 𝐻 e (x), and obtain the curriculum boundary 𝜅 (𝑛) from the running buffer 19: Update running hardness statistics, normalize 𝐻 e (x) )/𝜏c ) 20: Compute routing weights 𝛼 (x), 𝜔 (x) and curriculum weight 𝑤cur (x, 𝑛) = sigmoid( (𝜅 (𝑛) − 𝐻 21: Accumulate the sample contributions to Jplan (𝜃 ) and Jrend (𝜙 ) without resampling 22: end for 23: Update 𝜃 by text-GRPO using Jplan (𝜃 ) 24: Update 𝜙 by flow-GRPO using Jrend (𝜙 ) 15: 16:

directly into 3–5 binary verification questions per slot. The conversion follows four implementation constraints: slot exclusivity, so each question belongs to exactly one slot; cross-slot complementarity, so the same requirement is not repeated across slots; no cross-slot leakage, so one slot does not silently test another slot’s responsibility; and post-render verifiability, so every question can later be answered from the source image, the relevant slot text, and the realized edited image. In practice, 𝐶 mod targets the requested visible change, 𝐶 pre targets unchanged content, 𝐶 ovr targets global coherence and realism, and 𝐶 tip targets localized implementation details. This slot alignment allows the same checklist to support both module-specific rewards without changing the semantic target. During RL training, Gemini 3 Pro is no longer used. Qwen3-VL32B performs all online checklist-based scoring. On the planner side, it judges whether each rollout-time slot text was a useful plan for the observed edit under the matching checklist field. On the renderer side, it judges whether the final rendered image actually satisfies that same checklist field. The shared checklist, therefore, ties planner-side diagnosis and renderer-side execution scoring to the same pre-generated per-slot questions. The offline checklist generator also follows conservative default rules when the instruction under-specifies the edit. If the instruction is local, preservation questions default to protecting identity, pose, background, viewpoint, and other non-target content unless the instruction explicitly asks to modify them. If the instruction requires a background change, that requested change is moved into the Modify or Overall field and is not redundantly enforced as a

preservation constraint. Text-editing requests add exact renderedtext correctness to the relevant slot, while removal or inpainting requests add plausibility and local consistency checks so that filled regions remain compatible with surrounding texture, lighting, and geometry. The concrete prompt templates for these three stages are collected later in the template bank: planner rollout rewriting in Fig. S7, offline checklist generation in Fig. S8, and online scoring prompts in Figs. S9–S16.

2.3

Reward Instantiation in Practice

Question-level aggregation. For input x = (𝐼 src, 𝑐), plan e𝑖 , renplan dering y𝑖,𝑘 , slot 𝑗, and question 𝑞 ∈ 𝐶 ( 𝑗 ) , let 𝑎𝑖,𝑘,𝑞 ∈ {0, 1} denote the (𝑗) rend = I[Judge planner-side judgment and define 𝑎𝑖,𝑘,𝑞 rend (𝐼 src , y𝑖,𝑘 , 𝑐, 𝑒𝑖 , 𝑞) = Yes]. We encode Yes as 1 and No as 0. All questions within a slot are weighted equally, giving the per-render slot scores ∑︁ plan ∑︁ 1 1 (𝑗) (𝑗) rend 𝑠𝑖,𝑘 = (𝑗) 𝑎𝑖,𝑘,𝑞 , 𝑢𝑖,𝑘 = (𝑗) 𝑎𝑖,𝑘,𝑞 . (S21) |𝐶 | |𝐶 | (𝑗) (𝑗) 𝑞 ∈𝐶

𝑞 ∈𝐶

(𝑗) (𝑗) Thus, 𝑠𝑖,𝑘 and 𝑢𝑖,𝑘 are normalized Yes fractions in [0, 1], rather

than additional continuous outputs from the judge. Planner-side diagnostic. At scoring time, the shared checklist has already been generated offline by Gemini 3 Pro and loaded from cache, while the current planner policy has produced the rollouttime four-slot plan e𝑖 . Qwen3-VL-32B evaluates every question in each slot through single-question Yes/No judgments conditioned on the slot text, the original input, the realized edited image y𝑖,𝑘 , and

DARS for Instruction-Based Image Editing

Preprint.

(𝑗) the aligned checklist field 𝐶 ( 𝑗 ) . The resulting 𝑠𝑖,𝑘 is a post-render diagnostic of whether that slot functioned as a useful plan for the realized edit. The corresponding prompt templates are shown in (𝑗) (𝑗) Figs. S9–S12. Define the per-render gate 𝜎𝑖,𝑘 = I[𝑠𝑖,𝑘 ≥ 𝛿 𝑗 ]. The per-render planner reward is

𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C) =

4 ∑︁

(𝑗) 𝜆 𝑗 · 𝑠𝑖,𝑘 ·

𝑗=1

𝑗 −1 Ö

(𝑙 ) 𝜎𝑖,𝑘 .

(S22)

𝑙=1

We use (𝜆mod, 𝜆pre, 𝜆ovr, 𝜆tip ) = (0.4, 0.2, 0.2, 0.2) and 𝛿 𝑗 = 0.66 for every slot. After evaluating all 𝐾 renderings under plan e𝑖 , we compute 𝐾

plan

𝑟𝑖

=

1 ∑︁ 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C), 𝐾

𝐾

𝑠¯𝑖( 𝑗 ) =

𝑘=1

1 ∑︁ ( 𝑗 ) 𝑠𝑖,𝑘 . 𝐾

(S23)

𝑘=1

The first quantity yields the plan-level group-relative advantage, while the second and its analogous averaged-score gate 𝜎¯𝑖( 𝑗 ) = I[¯𝑠𝑖( 𝑗 ) ≥ 𝛿 𝑗 ] are used for the sign-aware token-level advantage in Section 1.1. Renderer-side execution scoring. The renderer reward changes the target of judgment while retaining the same checklist fields. For each 𝑗 and 𝑞 ∈ 𝐶 ( 𝑗 ) , Qwen3-VL-32B is conditioned on (𝐼 src, y𝑖,𝑘 , 𝑐, 𝑒𝑖( 𝑗 ) , 𝑞). The instruction and slot text specify the intended requirement; the judge scores whether the rendered image visibly satisfies it, not the (𝑗) quality of the slot text. The resulting 𝑢𝑖,𝑘 is the fraction of satisfied requirements in that slot. The corresponding prompt templates are shown in Figs. S13–S16. The per-render renderer reward is rend 𝑟𝑖,𝑘 = 𝑅rend (x, e𝑖 , y𝑖,𝑘 ; C) =

4 ∑︁

(𝑗) 𝛾 𝑗 𝑢𝑖,𝑘 .

(S24)

𝑗=1

Here, 𝛾 𝑗 controls the contribution of each slot to the final execution score. The 𝐾 renderer rewards under each fixed plan remain separate when computing the renderer-side group-relative advantage; they are not averaged before normalization. Shared reward reuse. The shared reward used for hardness estimation and variance decomposition is 𝑅share (x, e𝑖 , y𝑖,𝑘 ) = 𝜂 plan 𝑅plan (x, e𝑖 , y𝑖,𝑘 ; C)+𝜂 rend 𝑅rend (x, e𝑖 , y𝑖,𝑘 ; C). (S25) Its role is restricted to hardness estimation and variance decomposition; planner and renderer policy gradients still use their own module-specific rewards.

2.4

Illustrative Prefix-Gated Reward Example

We use a simple sequence-completion edit to illustrate how prefix gating changes planner-side credit assignment in practice. The source image 𝐼 src contains a diagonal chain of hexagons labeled “5, 7, 11, 13, 17, ?”, and the raw instruction 𝑐 is: “Replace the final question mark with the correct next number, while keeping the existing numbers, hexagons, and diagonal layout unchanged.” Figure S1 instantiates the same notation used in the main paper on this concrete case: Panels A–B define the input x = (𝐼 src, 𝑐), the structured ovr tip planner output e = (𝑒 mod, 𝑒 pre
, 𝑒 , 𝑒 ), and the shared checklist mod pre ovr tip C(x) = 𝐶 , 𝐶 , 𝐶 , 𝐶 ; Panels C–E then show three realized planner outcomes, namely Failure A, Failure B, and Success.

Consistent with the training definition in Eq. S21, each slot score is computed from its own checklist as 𝑠 ( 𝑗 ) = #Yes/|𝐶 ( 𝑗 ) |, with normalized slot weights (𝜆mod, 𝜆pre, 𝜆ovr, 𝜆tip ) = (0.4, 0.2, 0.2, 0.2) and a uniform activation threshold 𝛿 mod = 𝛿 pre = 𝛿 ovr = 𝛿 tip = 0.66. The planner-side reward for a realized outcome is then 𝑅plan = 𝜆mod𝑠 mod +𝜆pre𝑠 pre 𝜎mod +𝜆ovr𝑠 ovr 𝜎mod 𝜎pre +𝜆tip𝑠 tip 𝜎mod 𝜎pre 𝜎ovr, (S26) where 𝜎 𝑗 = I[𝑠 ( 𝑗 ) ≥ 𝛿 𝑗 ]. The question-level aggregation, slot weights, activation threshold, and prefix-gating rule are identical to those used in training. Failure A in Panel C shows the clearest advantage of prefix gating. The enhanced plan is locally neat in Preserve, Overall, and Tips, but the core Modify slot is wrong because 𝑒 mod proposes 21 instead of 19. The corresponding Modify score is therefore 𝑠 mod = 2/4 = 0.50 < 0.66, so 𝜎mod = 0. As a result, the prefix gate closes immediately after Modify, and the planner reward becomes (A) 𝑅plan = 0.4 · 0.50 + 0.2 · 1.00 · 0 + 0.2 · 0.75 · 0 + 0.2 · 1.00 · 0

= 0.20. (S27) Even though the rendered result still satisfies all five Preserve checks and most downstream checks, those later slots receive no additional credit once the causally prior Modify requirement fails. By contrast, an ungated weighted sum would still produce 0.75 because the strong downstream slots would partially compensate for the incorrect core reasoning. Failure B in Panel D illustrates a different error pattern. Here the Modify slot is correct, so 𝑠 mod = 4/4 = 1.00 and 𝜎mod = 1. However, the weak Preserve slot 𝑒 pre allows collateral changes near the target, and the rendered result changes the original “13” into “17” while also darkening the background. This drops the Preserve score to 𝑠 pre = 2/5 = 0.40 < 0.66, so 𝜎pre = 0. Thus, the reward still keeps the correct Modify credit, but blocks Overall and Tips from contributing: (B) 𝑅plan = 0.4 · 1.00 + 0.2 · 0.40 + 0.2 · 0.75 · 0 + 0.2 · 0.75 · 0

= 0.48.

(S28)

Again, the ungated weighted sum would be much higher (0.78) because the later slots would still receive credit despite the failed preservation constraint. Finally, Panel E shows the successful case. Here, all four slot texts align with the intended edit, and the realized image satisfies every checklist item across Modify, Preserve, Overall, and Tips. Therefore, all gates remain active, and the final planner reward is maximal: (success) 𝑅plan = 0.4 · 1.00 + 0.2 · 1.00 + 0.2 · 1.00 + 0.2 · 1.00

= 1.00.

(S29)

Taken together, the three cases in Fig. S1 make the intended creditassignment behavior explicit. Prefix gating does not simply average slot quality; instead, it follows the causal dependency structure of the four-slot plan. Modify must first specify the correct edit target, Preserve must then protect non-target content, and only after those upstream requirements are reliable, do Overall and Tips receive effective reward. This is precisely the behavior that the prefix-gated reward is designed to enforce during planner optimization.

Preprint.

Cao et al.

Panel A. Input image x and raw instruction

Panel B. Checklist C(x) and plan e Modify 𝐶 mod 1. Is the final “?” replaced with a numeral? 2. Is the new numeral exactly “19”? 3. Does the edit correctly continue the prime pattern after 17? 4. Is the target of the change the last hexagon? Preserve 𝐶 pre 1. Are 5 and 7 unchanged? 2. Is 11 unchanged? 3. Is 13 unchanged? 4. Are all non-target hexagons otherwise untouched? 5. Does the background remain white?

Raw instruction 𝑐: replace the final “?” with the correct next number, while keeping the existing numbers, hexagons, and diagonal layout unchanged. Ground-truth target: the sequence follows prime numbers, so the correct completion is “19”.

Overall 𝐶 ovr 1. Is the six-hexagon diagonal layout preserved? 2. Are size, spacing, and viewpoint consistent? 3. Does the final sequence remain globally coherent? 4. Is the result clean and artifact-free? Tips 𝐶 tip 1. Is only the last hexagon edited? 2. Does the new digit match the original font style? 3. Does the stroke/outline style remain consistent? 4. Is the new digit centered with similar spacing?

Structured plan e: e = (𝑒 mod, 𝑒 pre, 𝑒 ovr, 𝑒 tip ), where the four slots are Modify, Preserve, Overall, and Tips. Scoring rule: slot score 𝑠 ( 𝑗 ) = #Yes/|𝐶 ( 𝑗 ) | for 𝑗 ∈ {mod, pre, ovr, tip}. Gate rule: slot gate 𝜎 𝑗 = I[𝑠 ( 𝑗 ) ≥ 𝛿 𝑗 ] with 𝛿 mod = 𝛿 pre = 𝛿 ovr = 𝛿 tip = 0.66.

Panel C. Failure A: wrong Modify prompt

Panel D. Failure B: weak Preserve prompt

Panel E. Success: precise four-slot prompt

Modify 𝑒 mod : Put 21 in the last hexagon to continue the sequence. Preserve 𝑒 pre : Keep the first five numbered hexagons and the white background unchanged. Overall 𝑒 ovr : Preserve the diagonal layout and a smooth increasing pattern. Tips 𝑒 tip : Edit only the last hexagon and match the numeral style.

Modify 𝑒 mod : Put 19 in the last hexagon as the next prime after 17. Preserve 𝑒 pre : Keep the sequence readable, but nearby numerals or canvas tones may be adjusted if needed. Overall 𝑒 ovr : Preserve the diagonal layout and a plausible completed sequence. Tips 𝑒 tip : Harmonize the final region with neighboring hexagons so the ending looks locally consistent.

Modify 𝑒 mod : Put 19 in the last hexagon as the next prime after 17. Preserve 𝑒 pre : Keep 5, 7, 11, 13, 17, and the white background unchanged. Overall 𝑒 ovr : Preserve the prime progression, clean spacing, and original diagonal layout. Tips 𝑒 tip : Edit only the last hexagon and match the font, stroke, and centering.

Checklist hits: Modify 𝐶 mod : 2/4; Preserve 𝐶 pre : 5/5. Overall 𝐶 ovr : 3/4; Tips 𝐶 tip : 4/4. Scores: Modify 𝑠 mod = 0.50; Preserve 𝑠 pre = 1.00. Overall 𝑠 ovr = 0.75; Tips 𝑠 tip = 1.00. Gate: Modify gate 𝜎mod = 0 ⇒ Preserve, Overall, and Tips blocked. (A)

Reward: 𝑅plan = 0.20.

Checklist hits: Modify 𝐶 mod : 4/4; Preserve 𝐶 pre : 2/5. Overall 𝐶 ovr : 3/4; Tips 𝐶 tip : 3/4. Scores: Modify 𝑠 mod = 1.00; Preserve 𝑠 pre = 0.40. Overall 𝑠 ovr = 0.75; Tips 𝑠 tip = 0.75. Gate: Modify gate 𝜎mod = 1, Preserve gate 𝜎pre = 0 ⇒ Overall and Tips blocked.

Checklist hits: Modify 𝐶 mod : 4/4; Preserve 𝐶 pre : 5/5. Overall 𝐶 ovr : 4/4; Tips 𝐶 tip : 4/4. Scores: Modify 𝑠 mod = 1.00; Preserve 𝑠 pre = 1.00. Overall 𝑠 ovr = 1.00; Tips 𝑠 tip = 1.00. Gate: Modify, Preserve, and Overall gates all pass (𝜎mod = 𝜎pre = 𝜎ovr = 1) . (success)

Reward: 𝑅plan

= 1.00.

(B) Reward: 𝑅plan = 0.48.

Figure S1: Five-panel schematic of checklist-based prefix-gated planner reward on a sequence-completion edit. Panels A–B define the shared input, checklist, and structured plan; Panels C–E illustrate Failure A, Failure B, and Success under the same scoring and gating rules. Detailed interpretation is provided in Sec. 2.4.

DARS for Instruction-Based Image Editing

Preprint.

Table S1: Training recipe summary. Category

Hyperparameter

Value

VLM Policy

Base model Optimizer Learning rate Plans per input (𝑀 ) Total rollouts per input (𝑀 × 𝐾 ) Global update batch size KL penalty (𝛽 ) Max context / response length

Qwen3-VL-4B-Instruct AdamW 2 × 10 −6 4 16 256 1 × 10 −3 20,480 tokens / 4,096 tokens

Base model Fine-tuning method LoRA rank / scaling factor (𝛼 LoRA ) Optimizer Learning rate Weight decay Flow-GRPO clip range (𝜖r ) Renderings per plan (𝐾 ) SDE training timesteps EMA decay / update interval Max image resolution Inference steps (train / eval) Guidance scale

Qwen-Image-Edit-2511 LoRA 64 / 128 AdamW 2 × 10 −6 1 × 10 −4 1 × 10 −4 4 {1, 2, 3} 0.9 / 4 steps 1024 × 1024 10 / 40 4.0

Diffusion Policy

3.4

3.5

The main paper reports the shared training setup. Here we list only the policy-specific optimization settings needed to reproduce DARS; all shared quantities remain exactly as in the main text. The planner uses a standard AdamW text-policy update to rewrite the raw instruction online into a four-slot plan, while the renderer uses LoRA-based flow-GRPO with DiT-specific sampling controls. Relative to the controlled Joint RL baseline, DARS adds sharedrollout bookkeeping and slot-wise scoring without additional renderer calls.

3.2

Controlled Baseline and Evaluation Protocol

The controlled Joint RL + Adaptive Curriculum baseline matches DARS in backbones, training data, offline checklist cache, online reward judge, rollout budget, hardware, and adaptive curriculum. It differs only in replacing the structured four-slot planner and dual-level credit assignment with a free-form planner trace. All benchmark results are obtained from the official evaluation code and are independent of the training-time reward pipeline. Gemini 3 Pro and Qwen3-VL-32B are used only for checklist construction and training-time scoring, not for the final reported benchmark metrics. Across all in-house variants, we report the official overall benchmark scores, with PICA-Bench following its standard simple / detailed split.

3.3

KRIS-Bench Breakdown

Table S2 expands the main-paper KRIS-Bench result into its official factual-, conceptual-, and procedural-knowledge groups, so performance differences can be inspected at the subgroup level rather than only through the overall score. We also include the newly obtained fine-grained breakdown for Qwen-Image-Edit-2511 and update the best/second-best markings accordingly.

ImgEdit-Bench Breakdown

Table S4 expands the main-paper ImgEdit-Bench result into its nine official edit-operation groups, making it easier to see where different methods are strong or weak across common editing regimes rather than only in the aggregate. All entries follow the benchmark’s official GPT-4.1-based evaluation protocol.

3.6

GEdit-Bench-EN Breakdown

Table S5 expands the main-paper GEdit-Bench-EN result into its official split-wise metrics, so the relative behavior of different methods can be inspected separately on the Intersection subset and the Full set. All numbers come from the official evaluation code.

3.7 3 Experimental Setup Details 3.1 Training Setup and Hyperparameters

RISE-Bench Breakdown

Table S3 expands the main-paper RISE-Bench result into the four official reasoning categories of RISEBench-360: Temporal, Causal, Spatial, and Logical. This breakdown shows where gains come from across different reasoning types rather than only in the aggregate.

PICA-Bench Simple-Prompt Breakdown

Table S6 reports the fine-grained PICA-Bench results under the benchmark’s simple prompt setting, which we also refer to as the superficial prompt setting for clarity. We list all official physicalconsistency categories: LP, LSE, Reflection, Refraction, Deformation, Causality, GST, and LST, together with the overall score. As with the other supplementary benchmark tables, all values come from the official evaluation pipeline and are reported with two decimal places.

3.8

PICA-Bench Explicit-Prompt Breakdown

Table S7 reports the fine-grained PICA-Bench results under the benchmark’s detailed prompt setting, which corresponds to the explicit prompt protocol used in our evaluation pipeline. We keep the same category layout as in Table S6: LP, LSE, Reflection, Refraction, Deformation, Causality, GST, LST, and the overall score. All values come from the official evaluation pipeline and are reported with two decimal places.

3.9

Baseline Architecture Overview

Table S8 summarizes the main backbone composition and dominant control mechanism of each baseline based on the released model descriptions and our implementation notes.

4

Failure Attribution Validation Details

We sample 500 evaluation cases for the attribution study, with an even split between relatively simple and difficult edits (250 each), so the validation set covers both low-difficulty and high-difficulty failure modes. For each case, GPT-5 produces five independent failure-attribution labels, and the final case-level pseudo-label is the majority vote. The system prompt used for this annotation step is shown in Fig. S6. The GPT-5 annotator returns one of three raw labels: planner_side, renderer_side, or both. In the paper, these are reported respectively as planner-dominant, renderer-dominant, and mixed. After obtaining these case-level three-way pseudo-labels, we form three

Preprint.

Cao et al.

Table S2: Fine-grained KRIS-Bench results. We use the official GPT-4.1-based evaluation protocol. KRIS-Bench evaluates edited results with four dimensions: visual consistency, visual quality, instruction following, and knowledge plausibility, where the last dimension is used only for knowledge-intensive subsets such as Social Science, Natural Science, and Logical Reasoning. The raw judge outputs are defined on a 1–5 scale, while the benchmark reports normalized percentage scores with maximum 100. Higher is better for every metric. All values are reported with two decimal places. Factual Knowledge

Model

Conceptual Knowledge

Procedural Knowledge

Overall ↑

Attribute Perception ↑

Spatial Perception ↑

Temporal Perception ↑

Average Score ↑

Social Science ↑

Natural Science ↑

Average Score ↑

Logical Reasoning ↑

Instruction Decomposition ↑

Average Score ↑

64.03 61.10 67.85 69.67 64.03 68.18 67.52 73.67

55.17 66.75 64.83 64.33 68.08 67.75 74.42 71.75

56.31 58.45 79.17 80.18 71.28 87.50 84.91 81.53

59.57 62.42 69.04 70.21 66.13 71.27 71.92 74.54

66.00 66.20 58.65 63.50 66.75 68.85 67.25 71.85

63.19 63.04 57.39 60.62 67.26 66.49 68.10 67.95

63.86 63.80 57.69 61.32 67.13 67.06 67.89 68.89

38.04 53.08 51.38 49.75 50.50 51.63 48.00 56.04

64.83 54.61 67.29 62.67 53.89 64.94 71.28 68.22

49.52 53.74 58.24 55.29 51.95 57.33 57.98 61.26

59.57 61.02 61.24 62.58 63.26 66.04 66.78 68.80

74.75 50.00 75.67 84.33

77.70 74.77 82.43 85.36

71.60 58.26 74.33 80.75

61.75 53.45 71.40 82.95

58.48 56.22 70.59 81.20

59.27 55.55 70.79 81.62

56.21 49.58 56.63 71.00

81.83 57.10 86.50 89.39

67.19 52.79 69.43 78.89

64.42 55.72 71.54 80.72

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

68.82 56.88 71.67 78.21

Table S3: Fine-grained RISE-Bench results. We evaluate on RISEBench-360, the 360-sample version of the benchmark, using the official GPT-4.1-based judging pipeline. The benchmark scores each sample along instruction reasoning, appearance consistency, and visual plausibility, and counts a case as solved only when all applicable dimensions receive full marks. The table therefore reports Accuracy (%), i.e., the task success rate, for each reasoning category and for the overall set. Higher is better for every metric.

Method

Temporal ↑

Causal ↑

Spatial ↑

Logical ↑

Overall ↑

15.29 15.29 3.53 14.12 16.47 14.11 15.29 18.82

16.67 20.00 8.89 13.30 27.78 20.00 27.78 18.89

7.00 14.00 23.00 13.00 13.00 17.00 22.00 27.00

3.52 3.53 2.34 5.88 2.35 5.88 9.41 7.06

10.56 13.33 10.00 11.67 15.00 14.44 18.89 18.33

18.25 12.22 30.00 38.89

30.67 11.00 18.00 31.00

4.65 4.70 7.06 8.24

17.50 9.72 18.90 27.50

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

20.80 10.59 20.00 30.59

pairwise binary tasks by filtering the same 500-case pool into Planner vs. Renderer, Planner vs. Mixed, and Renderer vs. Mixed; we do not resample separately for the binary tasks, so the retained sample count in each setting is determined by the realized three-way label frequencies in this shared evaluation pool. Accuracy and Macro-F1 are computed within each resulting binary task, and AUROC is reported in the same one-vs.-one setting rather than under a three-way formulation. GPT-5 sees only the source image, the structured plan, and the edited image for each

sampled triplet; it does not see rollout statistics, checklist-based rewards, or slot scores.

5

Additional Qualitative Results

This section adds three complementary qualitative views: additional baseline comparisons, more reasoning-intensive edits, and more general editing cases.

DARS for Instruction-Based Image Editing

Preprint.

Table S4: Fine-grained ImgEdit-Bench results. The header enumerates the benchmark’s nine common editing tasks: Add, Remove, Adjust, Replace, Style, Background, Action, Hybrid, and Extract. ImgEdit-Bench evaluates each result with GPT-4.1 on a 1–5 scale from three perspectives: instruction following, editing quality, and detail preservation. Instruction following measures understanding of the prompt and target concept; editing quality measures how accurately the target region is manipulated; and detail preservation measures fidelity on content that should remain unchanged. Because instruction following is foundational and cannot be fully separated from the other two aspects, the editing-quality and detail-preservation scores are upper-bounded by the instruction-following score. The table reports the average score for each edit category, and higher is better for every metric. Method

Add ↑

Adjust ↑

Extract ↑

Replace ↑

Remove ↑

Background ↑

Style ↑

Hybrid ↑

Action ↑

Overall ↑

4.20 4.09 4.07 4.28 4.18 4.23 2.67 4.33

3.96 3.53 4.19 4.40 3.78 4.18 3.92 4.17

3.45 2.27 2.95 2.29 2.65 3.51 2.67 3.91

4.35 4.27 4.23 4.36 4.50 4.71 4.57 4.66

3.35 3.81 3.67 4.17 4.39 4.08 4.35 4.18

4.22 3.83 4.09 3.94 4.01 4.09 3.97 4.29

4.91 4.65 4.46 4.74 4.73 4.89 4.79 4.90

3.45 2.46 3.04 3.61 2.78 3.59 3.33 3.65

4.16 4.37 4.13 4.23 4.36 4.42 4.56 4.68

3.97 3.70 3.87 4.00 4.06 4.19 4.05 4.31

4.21 2.07 3.43 4.18

4.62 3.41 4.43 4.63

4.29 3.02 4.18 4.32

4.23 3.85 4.27 4.24

4.90 4.48 4.77 4.90

3.37 3.07 3.61 3.61

4.63 3.16 4.31 4.61

4.36 3.44 4.18 4.39

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

4.55 4.11 4.28 4.54

4.48 3.78 4.40 4.47

Table S5: Fine-grained GEdit-Bench-EN results. The header uses the official benchmark notation: G_SC denotes GPT-4.1-based semantic consistency, G_PQ denotes GPT-4.1-based perceptual quality, and G_O denotes the official overall score. Semantic consistency evaluates how well the edited result follows the instruction, while perceptual quality evaluates image naturalness and artifact level; both use a 0–10 scale. “Intersection subset” refers to the subset of test images for which all compared models return valid outputs, whereas “Full set” refers to the complete GEdit-Bench-EN test set. All values are reported with two decimal places.

Intersection subset

Method

Full set

G_SC ↑

G_PQ ↑

G_O ↑

G_SC ↑

G_PQ ↑

G_O ↑

7.49 7.31 7.32 8.32 7.18 6.84 8.08 8.20

7.80 6.96 7.60 7.89 7.21 8.21 7.76 7.84

7.22 6.67 6.90 7.73 6.62 6.63 7.54 7.64

7.43 7.10 7.15 8.05 7.06 6.70 7.93 8.12

7.76 6.89 7.58 7.85 7.20 8.16 7.66 7.88

7.16 6.47 6.74 7.47 6.52 6.48 7.37 7.59

7.68 7.66 7.97 8.01

7.24 6.55 7.79 7.98

7.65 6.87 8.08 8.49

7.38 7.62 7.96 8.00

6.97 6.43 7.59 7.86

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

7.67 7.03 8.33 8.57

Additional baseline comparisons. Figure S2 adds side-by-side comparisons on challenging edits involving temporal prediction, logical diagram reasoning, path planning, object removal, and symbolic

transformation. These cases make it easier to inspect where DARS preserves the target semantics more reliably than the omitted baselines.

Preprint.

Cao et al.

Instruction

Draw what it will look like Draw what it will look like ten after 30 seconds later. seconds in summer.

The caterpillar Draw what it wants to find the will look like shortest path to the leaf. Please mark after 30 the two shortest minutes. paths in the diagram.

Remove the water from the glass cup.

Move only two matchsticks on the left side of the equation to make the equation true. Draw the final equation.

Ori image ThinkGen

Step1X-Editv1p2

ThinkRL-Edit

UniREdit

Ours

Figure S2: Additional baseline comparisons on challenging edits. The figure adds side-by-side results for several baselines under the same evaluation setting. Instruction

This is a puzzle game. Draw the completed version of the puzzle in the image.

Draw the final state of the lever

Adjust all lunch boxes to the size of the largest lunch box

The image shows a maze. Please draw the solution path with a red line.

Ori image

Ours

Figure S3: Representative failure cases illustrating current limitations. From left to right, the examples highlight failure modes on puzzle completion, dense final-state reasoning in a mechanics-style diagram, precise relative-size control across multiple objects, and maze-path drawing. Together they show that the method remains weaker when the desired target is difficult to specify textually, when the edit depends on tightly coupled multi-step reasoning, or when exact geometric control is required.

DARS for Instruction-Based Image Editing

Preprint.

Table S6: Fine-grained PICA-Bench results under the simple-prompt setting. This setting corresponds to the superficial-prompt protocol in PICA-Bench and is evaluated by GPT-4.1 for instruction-based image editing models. All reported numbers are Accuracy (%). In the header, LP, LSE, GST, and LST denote Light Propagation, Light Source Effects, Global State Transition, and Local State Transition, respectively. Method

LP ↑

LSE ↑

Reflection ↑

Refraction ↑

Deformation ↑

Causality ↑

GST ↑

LST ↑

Overall ↑

64.39 61.99 56.04 63.52 62.15 60.96 63.10 63.05

66.34 66.17 52.24 58.12 63.20 56.68 69.26 59.98

64.51 67.50 59.27 66.61 62.73 61.18 67.34 65.75

65.46 54.38 41.35 62.44 51.35 41.16 51.96 53.35

50.06 53.11 50.42 53.41 52.55 56.92 58.31 55.34

51.26 52.90 45.60 48.76 48.84 46.07 51.47 50.76

68.38 65.68 52.91 52.60 60.21 59.53 65.79 64.50

66.34 56.83 45.67 51.89 54.76 52.66 61.51 60.98

60.82 60.40 51.02 56.15 57.30 55.13 61.48 59.64

71.86 49.29 70.86 72.04

62.33 48.88 59.55 64.62

57.93 52.33 53.71 64.37

50.70 43.82 53.06 53.59

67.89 61.26 62.90 65.50

61.34 52.15 63.38 66.16

63.20 52.00 61.90 64.19

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

63.89 52.99 66.86 62.52

69.91 51.29 66.19 67.85

Table S7: Fine-grained PICA-Bench results under the explicit-prompt setting. All reported numbers are Accuracy (%). All values are reported with two decimal places.

Method

LP ↑

LSE ↑

Reflection ↑

Refraction ↑

Deformation ↑

Causality ↑

GST ↑

LST ↑

Overall Score ↑

66.17 64.34 61.20 58.92 59.32 57.20 64.39 66.79

68.29 66.80 59.68 61.60 64.85 55.39 70.14 64.20

64.82 67.50 63.66 66.02 60.03 57.77 68.36 68.21

62.63 59.77 43.02 49.12 49.77 37.63 54.10 48.67

59.43 63.72 55.11 56.68 56.47 48.89 65.90 60.49

58.57 62.17 62.88 59.30 60.83 54.39 70.29 67.03

74.48 72.81 66.96 61.36 69.23 62.06 74.61 74.42

61.12 68.41 55.50 63.91 61.09 58.94 70.79 71.19

65.09 66.55 60.38 60.55 61.62 55.84 68.96 67.19

73.87 55.25 74.00 75.79

59.67 51.57 54.25 60.46

66.02 49.35 63.32 64.92

71.07 49.87 70.37 70.79

78.58 62.66 76.29 78.21

73.99 60.77 72.62 73.32

72.27 54.92 70.77 72.75

Other baselines ThinkGen UniREdit PromptRL Step1X-Edit-v1p2 UniReason 1.0 ThinkRL-Edit PhysicEdit EditThinker

Qwen-Image-Edit-2511-based methods Qwen-Image-Edit-2511 RePlan PromptEnhancerV2 DARS

67.36 46.84 67.46 67.79

75.82 56.27 74.01 74.97

Additional reasoning-intensive edits. Figure S4 shows that the advantage of DARS remains visible across several reasoning-heavy regimes, including rule-based game solving (e.g., rock-paper-scissors and tic-tac-toe), commonsense inference edits such as tip calculation and identifying the lightning rod as the protective facility to remove, and object completion or correction tasks involving missing or unreasonable structures. Additional general editing cases. Figure S5 shows that the benefit of DARS is not limited to explicitly puzzle-like instructions. Across more diverse general editing cases with richer object composition

and more complex inter-object relationships, DARS consistently demonstrates strong fine-grained control over local edits while preserving non-target content. These examples further suggest that the method maintains robust scene understanding and object-level recognition even in visually complex images, enabling precise edits without sacrificing overall consistency.

Preprint.

Cao et al.

Table S8: Detailed baseline architecture summary. Method ThinkRL-Edit ThinkGen Step1X-Edit-v1p2 RePlan EditThinker UniReason 1.0 PromptRL

5.1

Main Backbone

Key Structural Characteristics

Qwen3-VL-30B-A3B + Qwen-Image-Edit Qwen3-VL-8B-Think + OmniGen2-4B DiT

Checklist-style reward-driven RL with explicit CoT planning and reflection before image-to-image diffusion editing.

Qwen2.5-VL-7B + Step1X-Edit Qwen2.5-VL-7B + Qwen-Image-Edit-2511 Qwen3-VL-8B + Qwen-Image-Edit Bagel-14B Qwen2.5-VL-3B + FLUX .1-Kontext-12B

UniREdit

Bagel-14B

PhysicEdit

PhysicEdit data + Qwen-Image-Edit-2509

PromptEnhancerV2 Img2Img Edit

Qwen2.5-VL-32B + Qwen-Image-Edit-2511

Qwen-Image-Edit-2511

Qwen-Image-Edit-2511

Alternating GRPO between a reasoning VLM and a DiT generator. Multi-round reflection pipeline (up to five rounds) with coupled VLM and editor tuning. Region-aware decomposition: an RL-trained VLM predicts boxes and localized prompts for editing. Up to five rounds of reflection, with reasoning refinement concentrated on the VLM side. World-knowledge-aware prompt enhancement with reflection and consistency control. Joint RL over the prompt-side VLM and the image editor. Uses large-scale text reasoning data together with mixed real-world and game-world supervision for reasoning-oriented editing. Physically grounded reasoning with meta-query-based control for physics-aware image editing. Prompt-enhancement front-end paired with a strong external image editor; the public description mainly exposes the front-end/control structure rather than the full joint training recipe. Strong native end-to-end image editing model used as a general-purpose reference baseline.

Structured Plans for Qualitative Cases

Table S9 lists the four-slot decompositions for representative cases in Figs. S2–S5, so readers can directly inspect the semantic allocation across Modify, Preserve, Overall, and Tips.

6

Prompt Templates

We collect here the prompt templates used in three places in the pipeline. Figure S6 is used only in the GPT-5 failure-attribution validation study. Figure S7 is the planner-side four-slot rewriting prompt used during rollout to produce e. Figure S8 is the offline Gemini 3 Pro prompt used to generate the cached shared checklist C(x). During RL itself, Qwen3-VL-32B is used only for online singlequestion Yes/No scoring. Validation prompt. We first show the GPT-5 prompt used only for the failure-attribution validation study in Section 4. It is not part of the training or preprocessing pipeline. Planner rollout and checklist-generation prompts. We next show the two prompt families that create the semantic inputs to online scoring: the planner-side four-slot rewriting prompt used during rollout, and the Gemini 3 Pro prompt used before RL to generate the cached checklist target. Online reward-scoring prompts. After the planner produces a rollout-time four-slot plan and the checklist is loaded from the offline cache, reward evaluation proceeds through single-question prompts rather than a monolithic judgment. Each API call evaluates exactly one checklist question for exactly one slot, using Qwen3VL-32B as the online Yes/No judge during RL training. The slot text in these prompts comes from the current planner’s rollout-time

rewrite, while the checklist question comes from the offline Gemini 3 Pro cache. Planner-side prompts are post-render diagnostics: they ask whether the current slot text functioned as a good plan for the observed edit under one checklist item. Renderer-side prompts retain the raw instruction and relevant slot text as specification context but change the target: the judge scores whether the final rendered image visibly satisfies each checklist question, not whether the slot text is a good plan.

7

Limitations

Despite the gains reported in the main paper and the supplementary visualizations above, the current framework still has several clear limitations. One recurring difficulty arises when the desired output is hard to specify compactly in natural language and is more naturally described by a structured visual target. This includes tasks such as maze solving, puzzle completion, or other diagrammatic edits where the model must infer and render a precise final configuration rather than only apply a local semantic modification. A second limitation appears in cases that require either very dense chained reasoning or very precise geometric control. As shown in Fig. S3, representative failure modes include incomplete puzzle reconstruction, incorrect final-state reasoning in mechanicsstyle diagrams, inaccurate relative-size normalization across multiple objects, and unstable path drawing in mazes. These examples suggest that the current planner–renderer decomposition is still less reliable when the edit depends on long reasoning chains, diagramlevel state transitions, or exact multi-object scale relationships. Future progress will likely require stronger, structured intermediate representations, richer supervision for diagrammatic reasoning, and more explicit mechanisms for geometric control.

DARS for Instruction-Based Image Editing

Preprint.

Table S9: Actual four-slot plan decompositions for representative qualitative cases. Each row lists the raw user instruction together with its Modify, Preserve, Overall, and Tips fields in the same slot schema used throughout the paper. ID Raw

1

2

3

4

5

6

7

8

9

10

11

12

Modify

Preserve

Overall

Tips

Show the dry leaf catching fire under Preserve the surrounding green Add a charred, blackened texture Draw what it will The scene should depict the early the concentrated sunlight from the around the focal point and ensure the grass, the hand holding the look like after 30 stage of combustion in a visually magnifying glass, with a small flame magnifying glass, the position of smoke and flame originate naturally coherent and physically plausible seconds in from the concentrated light spot on and a plume of grey smoke rising from the lens, and the overall sunny summer. way under strong summer sunlight. the brown leaf. the bright focal point on the leaf. daylight illumination. Edit the image to depict the scene ten Preserve the background Maintain the original low camera The final image should look calm, angle, depth of field, and color grading, Draw what it will seconds later by removing the mountains, the tree-covered hills, smooth, and undisturbed, as if the look like ten suspended water droplets, the central the cloudy sky, the soft natural and ensure the reflections on the still splash event has fully settled into a seconds later. splash pillar, and the concentric ripple lighting, and the original lake water align naturally with the serene static scene. rings. setting. surrounding landscape. Change the scene to show the pot after the milk has been boiling for 30 Preserve the shape and material of Add bubbly and browned residue The result should convey a realistic Draw what it will minutes, with messy frothy milk spills the pot, the visible gas flame, the along the inner rim, mix darker smoke long-boiling, overheating situation with the steam, and make the spilled look like after 30 down the black pot and onto the gas stove grate structure, the blurred with coherent interactions between minutes. stove grate, along with a reduced, background, and the original warm milk follow gravity and pool plausibly milk, pot, stove, steam, and smoke. thickened, overheated surface inside lighting and perspective. on the stove surfaces. the pot. The caterpillar wants to find the Mark the two shortest surface paths Draw one path across the front-left Preserve the original blue cube, the The final image should look like a face and top face, and the other across shortest path to from the orange caterpillar at the green leaf, the orange caterpillar, clean diagrammatic solution that the leaf. Please bottom front corner of the cube to the the front-right face and top face, and the white background exactly remains geometrically consistent mark the two green leaf on the top face using two rendering both lines clearly and as they are. with the cube’s perspective. aligned to the cube surface geometry. shortest paths in distinct overlaid lines. the diagram. Move only two Turn the 8 into a 3 by removing the matchsticks on the Transform the left side of the Preserve the minus sign, the equals The final image must show a two left vertical matches, turn the 7 left side of the matchstick equation so the final result sign, the digit 0 on the right, the coherent, mathematically correct, into a 3 by adding a middle and equation to make reads 3 - 3 = 0 by changing the two left white background, and the overall and well-aligned matchstick bottom horizontal match, and match the equation true. digits accordingly. matchstick style. equation. the existing dark match heads and Draw the final rectangular shafts. equation. Eliminate optical refraction, Remove the water from the glass cup so Preserve the “HEAD ATP” text, the The final result should look like a magnification, and distortion from the Remove the water the glass interior appears empty and fuzzy tennis-ball texture, the shape realistic dry glass container holding ball, while maintaining the original from the glass cup. dry, and the tennis ball appears without and material of the clear glass cup, the ball naturally at the bottom lighting, shadows, and camera liquid-induced distortion. and the bright cyan background. under the same viewing conditions. viewpoint. This is a rockPreserve the top-left hand showing paper-scissors Redraw the bottom hand as a fist with The final image should remain a the index and middle fingers extended game. Change one Change the bottom hand gesture from scissors and the right hand showing consistent rock-paper-scissors in a V-shape, matching the original hand gesture so paper to scissors so that the rock player rock exactly as they are, along with illustration with correct game logic wins the game. the black-and-white cartoon line thickness and glove-like cartoon that the player and unchanged visual composition. design. who played rock drawing style. wins. Preserve the blue grid lines, the Position the new red O in the Place the “O” in The final image should look like a Add a red O mark to the top center blue X marks, the existing red O top-center cell, match the existing red the appropriate consistent hand-drawn tic-tac-toe square of the tic-tac-toe grid to marks, the lined notebook paper ink color and rough sketch style, and spot on the board where the O player wins background, and the pause icon in keep the stroke thickness consistent tic-tac-toe board complete the winning row. naturally. the top-right corner. with the other O marks. to win the game. The receipt should still look like a Format the new line using the same Add a 10% tip to Add a new line below the total amount Preserve all existing receipt text, coherent and authentic printed font, alignment, spacing, and the receipt in the of $45.00 labeled “Tip:” followed by layout, spacing, and visual style. receipt with the new tip line typographic style as the rest of the image. “$4.50”. integrated naturally. receipt. Preserve the red roof tiles, the Seamlessly reconstruct the roof ridge Remove the lightning rod structure The final image should look like a Remove the concrete building structure, the and fill the removed area with from the peak of the red tiled roof, realistic roof scene with no visible facility used to windows, the large satellite dish on surrounding sky so the roofline and including the vertical pole and its trace of the removed protection protect the house. the right, and the original sky gradient remain continuous and spherical components. facility. daylighting and perspective. undisturbed. Add the missing top- and right-face The completed puzzle should look Preserve all existing blue, orange, blocks with white stickers on Fill in the missing cubies to restore the like a coherent, fully assembled Complete the and white stickers, the black plastic top-facing surfaces and orange puzzle into a complete 4x4x4 cube with cube with consistent structure, Rubik’s Cube. body, and the purple patterned stickers on right-facing surfaces, solid cubic geometry. perspective, and material background. aligning them perfectly with the appearance. existing grid lines and shading. Use a recognized outlet configuration Correct the unrealistic and scattered Correct Preserve the beige square faceplate, The final socket should look like a such as a standard 3-pin or universal arrangement of holes on the wall unreasonable parts the outer frame, the light gray realistic manufactured electrical socket design, organize the holes into socket by replacing them with a of the socket in speckled wall background, and the outlet with a clean, balanced, and logical pairs or groups, and render standard, symmetrical, and functional the image. overall lighting and texture. logically organized layout. their depth, shading, and geometry outlet pattern centered on the plate. with precise plastic-like realism.

Preprint.

Cao et al.

Instruction

This is a rockpaper-scissors game. Change one hand gesture so that the player who played rock wins.

Place the ‘O’ in the appropriate spot on the tictac-toe board to win the game.

Add a 10% tip to the receipt in the image.

Remove the facility used to protect the house.

Complete the Rubik’s Cube.

Correct unreasonable parts of the socket in the image.

Ori image Qwen-ImageEdit-2511

ThinkGen

UniREdit

PromptRL

Step1X-Editv1p2

UniReason

ThinkRL-Edit

PhysicEdit

EditThinker

RePlan

Prompt EnhancerV2

Ours

Figure S4: Additional reasoning-intensive editing examples. The figure covers temporal, spatial, logical, physical, and knowledgeintensive editing cases.

DARS for Instruction-Based Image Editing

Instruction

Raise the person's right arm.

Preprint.

Change the military vehicle in the picture to be set in a beach environment.

Extract the Change the architectural palm trees and structures visible in sandy beach in the background of the background the image, to a snowy separating them landscape with cleanly from the sky pine trees. and surrounding environment.

Extract the footprint in the sand from the Image.

Remove the phone from the person’s hand.

Replace the person in the mirror with a wardrobe.

Ori image Qwen-ImageEdit-2511

ThinkGen

UniREdit

PromptRL

Step1X-Editv1p2

UniReason

ThinkRL-Edit

PhysicEdit

EditThinker

RePlan

Prompt EnhancerV2

Ours

Figure S5: Additional general editing examples. The figure covers localized modifications, preservation-sensitive edits, semantic scene changes, extraction, and removal cases.

Preprint.

Cao et al.

System Prompt for GPT-5 Failure Attribution You are a strict failure-attribution evaluator for instruction-based image editing. You will be given: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction - Enhanced Edit Instruction Your task is to determine whether the main failure comes from: 1. planner_side 2. renderer_side 3. both Definitions: planner_side: Choose this when the enhanced instruction is itself inadequate as an editing plan. This includes cases where the enhanced instruction is vague, incomplete, mislocalized, misbound to the wrong target, missing necessary preservation constraints, missing necessary global consistency constraints, missing necessary local implementation details, or otherwise fails to faithfully and operationally specify the user’s intent for the given source image. In this case, the rendered failure can be reasonably explained primarily by planning/instruction deficiency. renderer_side: Choose this when the enhanced instruction is already a sufficient and reasonable editing plan for the given source image, but the final rendered image fails to execute it correctly. This includes cases where the rendered result misses the requested edit, applies it to the wrong target, causes unnecessary collateral changes, breaks realism or consistency, or otherwise fails despite the enhanced instruction being adequate. both: Choose this when both sides contribute materially to the failure. This includes cases where the enhanced instruction is meaningfully flawed or insufficient, and the final rendered result also shows execution failures that cannot be explained only by the planning issue. Use both when there is substantial planner deficiency and substantial renderer deficiency at the same time. Core evaluation principle: You must separate: - plan adequacy: whether the enhanced instruction is a faithful, grounded, and operational plan for the task, and - execution adequacy: whether the final rendered image successfully follows that plan. Decision procedure: 1. Compare the original user instruction with the enhanced instruction. 2. Judge whether the enhanced instruction is a good plan for editing the source image. 3. Compare the final rendered image against the source image and the enhanced instruction. 4. Determine whether the observed failure is mainly due to bad planning, bad execution, or both. Important rules: - Judge the enhanced instruction relative to the source image, not in the abstract. - Do not prefer planner_side merely because the enhanced instruction could be even better; it must be materially inadequate. - Do not prefer renderer_side if the enhanced instruction is clearly under-specified or misleading in a way that would likely cause the failure. - Use both only when both deficiencies are real and important. - If the enhanced instruction is adequate and the image still fails, choose renderer_side. - If the enhanced instruction is inadequate and the image failure is consistent with that planning failure, choose planner_side. - If the enhanced instruction is inadequate and the rendered image also introduces additional execution failures beyond that, choose both. Examples of planner-side evidence: - wrong target identified - missing preservation constraint for important unchanged content - missing global consistency constraint when clearly needed - ambiguous local instruction that leaves critical placement or binding unspecified - rewritten instruction drifts from user intent Examples of renderer-side evidence: - requested edit missing or weak despite adequate instruction - correct target not edited - wrong region edited - identity, pose, background, or unrelated content altered unnecessarily - geometry, anatomy, lighting, perspective, or local realism broken despite adequate instruction Output requirement: Return exactly one label and nothing else. Allowed outputs: planner_side renderer_side both

Figure S6: System prompt used for GPT-5 failure-attribution annotation in the routing-signal validation study. The raw outputs planner_side, renderer_side, and both are mapped in the paper to planner-dominant, renderer-dominant, and mixed, respectively.

DARS for Instruction-Based Image Editing

Preprint.

System Prompt for Planner Rollout Four-Slot Rewriting You are an expert in writing prompts for instruction-based image editing. Given a source image and a user instruction, rewrite the instruction into a detailed, precise prompt for an image editing model. Organize the rewritten prompt into four explicitly tagged fields that specify what should change, what must be preserved, where the edit applies, and how the result should remain consistent with the original image. ### I. Editing Objectives The rewritten prompt must faithfully reflect the user’s requested modification while preserving all unrelated content. 1. Edit Specificity: Clearly identify the target of the edit: object, person, attribute, region, text, background, style, pose, action, or layout. 2. Preservation of Unedited Content: Explicitly preserve all content not mentioned in the user’s instruction, including identity, composition, viewpoint, background, lighting, color harmony, structure, and scene context, unless the instruction requires changing them. 3. Localized and Grounded Editing: Describe the edit with precise grounding in the image, including spatial position, entity relationships, and attribute bindings. 4. Structural and Visual Consistency: Ensure the final result remains realistic and coherent in geometry, anatomy, perspective, shadows, reflections, texture, and occlusion. 5. Minimal Necessary Change: If the instruction is ambiguous, prefer the smallest valid edit that satisfies the user’s intent without introducing unnecessary changes. ### II. Tagged Field Structure The rewritten prompt must contain exactly four tagged fields, with one field for each slot. Content within each field may use free-form natural language. 1. Modify: State the main requested edit directly and clearly. 2. Preserve: Explicitly state the important elements that should remain unchanged. 3. Overall: Describe the desired image-level coherence, including consistency in lighting, perspective, identity, style, and scene continuity. 4. Tips: Provide targeted implementation details about the edited subject or region, including attributes, actions, position, relations, and appearance. ### III. Grammatical Rules 1. Use clear, descriptive, and objective language. 2. Use present tense consistently. 3. Use precise spatial and relational expressions. 4. Use specific adjectives only when supported by the source image or user instruction. 5. Avoid vague expressions such as “make it better” unless the user explicitly requests them. 6. Prefer direct description over meta commentary. 7. Keep the rewritten prompt compact but sufficiently detailed for high-quality editing. ### IV. Editing-Specific Rules 1. Only provide the final rewritten editing prompt. Do not use markdown format. 2. Do not explain your reasoning. 3. Do not mention “source image”, “input image”, “edited image”, “before”, or “after” unless necessary for disambiguation. 4. Do not add new objects, text, attributes, events, or background elements that are not required by the user’s instruction. 5. If the edit concerns only one entity among multiple entities, preserve the others and bind the edit to the correct target. 6. If the instruction changes a local attribute, preserve identity, pose, and other attributes unless explicitly asked otherwise. 7. If the instruction changes the background, preserve the main subject unless explicitly asked otherwise. 8. If the instruction removes an object, ensure the filled region is visually plausible and consistent with the surrounding scene. 9. If the instruction inserts or edits text, preserve the exact requested wording and format it as “rendered text”. 10. If the instruction does not request a style change, preserve the original image style, medium, camera viewpoint, and overall tone. 11. If the instruction requests a style transfer, preserve the scene semantics, object layout, and key subject identity unless explicitly asked otherwise. 12. If human hands, faces, bodies, or animal limbs are involved, ensure structural integrity and anatomical plausibility. 13. If reflections, mirrors, glass, water, or shadows are present, keep them physically consistent with the requested edit. 14. Output exactly four tagged fields, each appearing once, in the order Modify, Preserve, Overall, Tips. <Modify>...</Modify> <Preserve>...</Preserve> <Overall>...</Overall> <Tips>...</Tips> 15. Do not output any text outside these tags. Next, I will provide: - Source image: - User edit instruction: Please provide the rewritten four-slot editing prompt using the required tags:

Figure S7: Planner-side prompt template used during rollout to rewrite a raw editing instruction into a four-slot plan. The resulting e is the planner output used by the renderer and by planner-side reward scoring.

Preprint.

Cao et al.

System Prompt for Shared Checklist Generation You are an expert evaluator for instruction-based image editing. Your task is to generate a single shared slot-aligned checklist directly from the source image and user instruction. The four slots are: 1. Modify 2. Preserve 3. Overall 4. Tips You will be given: - a source image, - and a user edit instruction. You must generate four lists of binary Yes/No verification questions: - modify - preserve - overall - tips ### I. Design Constraints A. Slot exclusivity: Each question must belong to exactly one slot. Do not write a question that simultaneously checks multiple slots. B. Cross-slot complementarity: The four lists must be complementary rather than redundant. Do not repeat the same semantic requirement across different slots using different wording. C. No cross-slot interference: A question in one slot must not depend on requirements that should be judged in another slot. Modify checks the requested change, Preserve checks what remains unchanged, Overall checks global coherence, and Tips checks targeted local implementation details. D. Post-render verifiability: Every question must be answerable later by looking at the source image, the relevant slot text, and the final rendered image. E. Strict binary evaluability: Each question must be clear, objective, and answerable with Yes or No. “Yes” must always mean success. F. Minimal-change principle: If the instruction is ambiguous, prefer the smallest valid edit and avoid inventing unnecessary requirements. ### II. Slot Definitions Modify: the main requested edit, target binding, and intended visible change. Preserve: important content that should remain unchanged. Overall: image-level coherence, realism, perspective, identity/style continuity, structural consistency, and physical plausibility when relevant. Tips: targeted local implementation details such as exact region, relative position, attribute binding, text rendering, fill plausibility, or local appearance constraints when relevant. ### III. Checklist Construction Requirements 1. Generate 3 to 5 questions per slot. 2. Questions must be concise. 3. Questions must be non-redundant within each slot. 4. Questions must be non-redundant across slots. 5. Do not include generic writing-quality questions. 6. Do not ask about hidden reasoning or internal model behavior. 7. Do not ask about anything that cannot be visually checked after rendering. ### IV. Output Format Output ONLY a valid JSON object; three-question example: { "modify": ["Question 1?", "Question 2?", "Question 3?"], "preserve": ["Question 1?", "Question 2?", "Question 3?"], "overall": ["Question 1?", "Question 2?", "Question 3?"], "tips": ["Question 1?", "Question 2?", "Question 3?"] } Do not output markdown, explanations, notes, comments, or any extra text.

Figure S8: Prompt template used in the offline Gemini 3 Pro preprocessing stage to generate the shared slot-aligned checklist C(x) from the source image and user instruction. The template enforces slot exclusivity, cross-slot complementarity, post-render verifiability, and strict binary Yes/No evaluability so that the same pre-generated checklist can be reused across planner reward, renderer reward, hardness estimation, and routing.

DARS for Instruction-Based Image Editing

Preprint.

System Prompt for Planner-Side Modify Evaluation You are a strict planner-side evaluator for the Modify slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the current Modify slot text functions as a good planning statement for the requested edit, after considering the final rendered result. Evaluation target: - Judge the quality of the current Modify slot as a plan. - Use the final rendered image only as post-render evidence showing whether the slot was sufficiently specific, correctly bound, and operational. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Do not reward generic wording. - Answer “No” if the slot is vague, misbound, incomplete, or if the rendered result reveals that this Modify slot failed to specify the needed edit clearly enough. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Planner-Side Modify Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Current Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Modify-slot checklist question. Decision rule: Answer Yes only if the Modify slot text clearly plans the requested visible change on the correct target, and the final rendered image does not reveal a planning failure for this requirement. Answer No if: - the slot text is vague, - the target is misidentified, - the edit is weakly specified, - the slot omits a necessary binding, - or the rendered result reveals that this Modify plan was inadequate. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S9: Planner-side single-question prompt for the Modify slot. The prompt evaluates one checklist question as a post-render diagnostic of whether the current Modify slot text functioned as an adequate plan.

Preprint.

Cao et al.

System Prompt for Planner-Side Preserve Evaluation You are a strict planner-side evaluator for the Preserve slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the current Preserve slot text functions as a good preservation plan, after considering the final rendered result. Evaluation target: - Judge whether the Preserve slot explicitly and adequately protects important content that should remain unchanged. - Use the final rendered image as post-render evidence of whether the preservation constraints were sufficient. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Answer “No” if the Preserve slot fails to explicitly protect important unchanged content, or if the rendered result reveals collateral changes that this slot should have constrained. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Planner-Side Preserve Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Current Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Preserve-slot checklist question. Decision rule: Answer Yes only if the Preserve slot text explicitly protects the relevant unchanged content, and the final rendered image does not reveal a preservation failure for this requirement. Answer No if: - important unchanged content is not explicitly protected, - the preservation scope is too weak or incomplete, - the slot leaves room for unnecessary collateral edits, - or the rendered result reveals preservation failures that this slot should have prevented. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S10: Planner-side single-question prompt for the Preserve slot. The prompt evaluates one checklist question as a post-render diagnostic of whether the current Preserve slot text adequately protected unchanged content.

DARS for Instruction-Based Image Editing

Preprint.

System Prompt for Planner-Side Overall Evaluation You are a strict planner-side evaluator for the Overall slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the current Overall slot text functions as a good global-coherence plan, after considering the final rendered result. Evaluation target: - Judge whether the Overall slot meaningfully constrains image-level coherence, realism, continuity, and consistency. - Use the final rendered image as post-render evidence showing whether the overall constraints were adequate. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Answer “No” if the Overall slot is generic, weak, non-operational, or if the rendered result reveals global inconsistencies that this slot should have constrained. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Planner-Side Overall Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Current Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Overall-slot checklist question. Decision rule: Answer Yes only if the Overall slot text meaningfully plans global coherence and consistency, and the final rendered image does not expose a failure that this slot should have constrained. Answer No if: - the slot is too generic, - the global consistency requirement is weakly specified, - the slot fails to constrain realism, continuity, or coherence where needed, - or the rendered result reveals global inconsistency that this slot should have helped prevent. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S11: Planner-side single-question prompt for the Overall slot. The prompt evaluates one checklist question as a post-render diagnostic of whether the current Overall slot text adequately constrained global coherence.

Preprint.

Cao et al.

System Prompt for Planner-Side Tips Evaluation You are a strict planner-side evaluator for the Tips slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the current Tips slot text functions as a good targeted implementation plan, after considering the final rendered result. Evaluation target: - Judge whether the Tips slot supplies useful local, grounded, implementation-critical details. - Use the final rendered image as post-render evidence showing whether these local details were necessary and adequately specified. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Answer “No” if the Tips slot is vague, generic, ungrounded, or if the rendered result reveals local failures that better Tips guidance should have constrained. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Planner-Side Tips Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Current Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Tips-slot checklist question. Decision rule: Answer Yes only if the Tips slot text provides the local grounded detail needed for this requirement, and the final rendered image does not reveal that such detail was missing or inadequate. Answer No if: - the tip is vague or generic, - the local target or relation is not grounded, - the implementation detail is missing, - the slot does not bind the detail to the correct region or entity, - or the rendered result reveals a local failure that better Tips guidance should have prevented. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S12: Planner-side single-question prompt for the Tips slot. The prompt evaluates one checklist question as a post-render diagnostic of whether the current Tips slot text provided adequate local implementation guidance.

DARS for Instruction-Based Image Editing

Preprint.

System Prompt for Renderer-Side Modify Evaluation You are a strict renderer-side evaluator for the Modify slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the final rendered image successfully executes the required Modify-slot condition. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Judge visible evidence only. - Use the source image as the reference image. - Use the user instruction and slot text only as the intended specification. - Answer “No” if the requested change is missing, weak, misplaced, incomplete, bound to the wrong target, or visually ambiguous. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Renderer-Side Modify Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Relevant Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Modify-slot checklist question from visible evidence. Decision rule: Answer Yes only if Picture 2 clearly executes the requested visible modification specified by this requirement. Answer No if: - the edit is absent, - too weak, - applied to the wrong target, - applied in the wrong way, - or visually ambiguous. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S13: Renderer-side single-question prompt for the Modify slot. The prompt evaluates one checklist question as an executionscoring query over visible modification quality.

Preprint.

Cao et al.

System Prompt for Renderer-Side Preserve Evaluation You are a strict renderer-side evaluator for the Preserve slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the final rendered image successfully preserves the required unchanged content. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Judge visible evidence only. - Use the source image as the preservation reference. - Answer “No” if there are unnecessary changes to content that should have remained unchanged, even if the main edit is otherwise correct. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Renderer-Side Preserve Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Relevant Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Preserve-slot checklist question from visible evidence. Decision rule: Answer Yes only if the relevant content remains unchanged in Picture 2 relative to Picture 1. Answer No if: - identity changes unnecessarily, - unrelated objects or regions are altered, - composition, viewpoint, or background shifts without need, - or other collateral changes violate this preservation requirement. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S14: Renderer-side single-question prompt for the Preserve slot. The prompt evaluates one checklist question as an executionscoring query over whether unchanged content is actually preserved.

DARS for Instruction-Based Image Editing

Preprint.

System Prompt for Renderer-Side Overall Evaluation You are a strict renderer-side evaluator for the Overall slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the final rendered image satisfies the required global coherence and realism condition. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Judge visible evidence only. - Answer “No” if geometry, anatomy, perspective, lighting, texture continuity, style continuity, shadows, reflections, or scene logic fail in a way relevant to this question. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Renderer-Side Overall Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Relevant Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Overall-slot checklist question from visible evidence. Decision rule: Answer Yes only if Picture 2 remains globally coherent and realistic for this requirement. Answer No if: - perspective breaks, - lighting is inconsistent, - identity or style continuity fails where relevant, - anatomy or geometry is implausible, - or overall scene continuity is broken. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S15: Renderer-side single-question prompt for the Overall slot. The prompt evaluates one checklist question as an executionscoring query over global coherence and realism.

Preprint.

Cao et al.

System Prompt for Renderer-Side Tips Evaluation You are a strict renderer-side evaluator for the Tips slot in instruction-based image editing. You will judge exactly one checklist question in this API call. Your task is to determine whether the final rendered image satisfies the required local implementation detail. Rules: - Judge only whether the current question is satisfied. - Do not apply any criterion that is not explicitly required by this question. - Do not turn this question into an overall evaluation of the whole result. - Judge visible evidence only. - Answer “No” if the local detail is missing, weak, misplaced, bound to the wrong entity, visually inconsistent, or only partially satisfied. Output ONLY a valid JSON object. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." } Do not output anything else. User Prompt for Renderer-Side Tips Evaluation Input: - Source Image: Picture 1 - Final Edited Image: Picture 2 - User Edit Instruction: {user_instruction} - Relevant Slot Text: {slot_text} - Current Checklist Question: {question_text} Task: Judge only this one Tips-slot checklist question from visible evidence. Decision rule: Answer Yes only if Picture 2 satisfies the required local implementation detail for this question. Answer No if: - the local target is wrong, - the position or relation is wrong, - the attribute binding is wrong, - rendered text is incorrect, - local appearance is inconsistent, - or fill/inpainting behavior is not plausible when relevant. Output JSON only. Example for a Yes judgment: { "answer": "Yes", "brief_reason": "Short factual reason." }

Figure S16: Renderer-side single-question prompt for the Tips slot. The prompt evaluates one checklist question as an executionscoring query over local implementation quality.

Related documents

• Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion#1050403
• Scaffolded AI-Supported Problem-Based Learning for Medical Interns: Exploratory Retrospectively Registered Randomized Controlled Evaluation With a Voluntary Feasibility Follow-Up.#653305
• Augmenting patient safety surveillance in radiation oncology with large language model-based root cause analysis: A proof-of-concept study#1092083
• AI-Assisted Social Story Intervention for Special Education: The Design of AdaptED Stories#1050428
• An overview of the Navy Automated Transportation Documentation System (NAVADS)#497195