arXiv CSopen access

Paint-Anything: Unified Any-Color Control for Image Generation and Editing

Paint-Anything: Unified Any-Color Control for Image Generation and Editing Ji Xie1,2 , Dewei Zhou1,2 , Xinyu Huang1 , Zhennan Chen1,3 , Xun Wang1

arXiv:2609.20816v1 [cs.CV] 17 Sep 2026

1

ByteDance Seed, 2 Zhejiang University, 3 Nanjing University

Abstract Professional design requires any-color control: the ability to specify an object’s target color with any 24-bit hex value for image generation and editing. Prior work has explored color generation, editing, and colorization, but often relies on dedicated color representations or specialized inference procedures. Advances in large language models offer a simpler starting point: even compact models can associate hex values with color semantics. We present Paint-Anything, which learns a shared hex-prompt interface for generation and editing through object-level color supervision. We develop a data pipeline that constructs Paint-500K from real images through object grounding, perceptual color labeling, and editing-pair synthesis. Since shadows make real-image labels only approximate colors, we complement this supervision with pure-color anchors whose pixels exactly match their paired hex values. These anchors are used only at high-noise timesteps, leaving low-noise training to natural images. We further introduce Any Color Benchmark (ACBench), comprising ACBench-T2I and ACBench-Edit, to measure object-level hex color fidelity across both tasks. On FLUX.2-4B, Paint-Anything improves ACBench-T2I and ACBench-Edit scores by 85.3% and 28.3%, respectively, relative to the base model, with ablations supporting the training recipe. It also achieves the highest average CompColor score among the compared methods. Date: September 18, 2026

1

Introduction

How far is AI from a designer? Today’s image models can imagine rich scenes and render them in remarkable

detail [25, 27, 46, 60, 61, 72, 74, 75]. Yet professional design requires more than a convincing image: it requires following a color specification. A brand designer needs a logo in the brand’s signature blue; an online retailer needs product images in the specified red of a new collection. Neither requirement is captured by asking for “blue” or “red”: both call for explicit hex values. Users should be able to choose an object’s color as directly as a painter chooses colors for a canvas. We call this capability any-color control: users specify target colors through 24-bit hex values. Prior work has explored various approaches to color control. Some methods introduce task-specific modules or learned color tokens, while others use training-free inference-time techniques such as sampling guidance or attention/value manipulation [3, 11, 39, 45, 48, 55, 56, 69]. These designs can be difficult to scale across colors, hard to transfer to new architectures, or costly at inference. As a result, any-color generation, colorization, and editing are often treated as separate problems rather than as one native prompt-following capability.

1

set the lamp’s color to

#10A9FF

I’m giving you a palette: #72B9F3, #D5DFE6, #B1B1A8, #788694, #58626F, #323D4E, please generate the image: Mount Fuji, oil painting.

#40F5BB

#BD9DDB

#F9E9A2

#FF73B0

A #A7C7E7 bed, lit by two warm bedside lamps, one A woman wears a #D8A7B1 and #EED6D3 checked #FFD580, one #F6EFDA, and a subtle #7DF9FF kimono, gemstone earrings in #95B5B7, #C26D3A neon light casting volumetric glow, moody lighting. flat woven hat, illuminated by #C8CBD9 light.

Figure 1 One hex-prompt interface for generation and editing. A single finetuned model supports object recoloring and hex-conditioned generation through explicit color values in the prompt.

Fortunately, advances in large language models make hexadecimal color prompting a promising interface for this goal. A 24-bit RGB value can be written directly in text and bound to an object phrase, making it possible for a single model to support different any-color generation tasks with the same prompt language. As a simple motivating probe, Figure 2 shows that even a small LLM, Qwen3-4B [64], can parse raw hex strings such as #F0FFF1 and #E8D00A into plausible color descriptions. This suggests that hex strings can be used as a prompt-native numeric color interface.

? Could you describe the color #E8D00A ? #F0FFF1

LLM Qwen3-4B

Certainly! The color #F0FFF1 is a light, pale shade that is often described as a soft, almost translucent greenish-white..... Certainly! The color #E8D00A is a shade of yellow. Here's a detailed description: ...... It is similar to gold, but with a more muted tone, and it's a bit more orangey than a typical yellow.

Figure 2 Hex color strings are legible to modern language models. Even a compact 4B language model such as Qwen3-4B can associate raw hex codes such as #F0FFF1 and #E8D00A with plausible color semantics, suggesting that hex strings can serve as a unified text interface for any-color generation and editing.

In this work, we present Paint-Anything, a unified model for any-color controllable generation and editing. Paint-Anything keeps color specification entirely inside the text prompt using explicit spans such as <color> #AABBCC </color>, so that one interface covers both generation and editing. Instead of adding a separate inference-time color controller, we construct Paint-500K through a filtered data pipeline that turns real-world images into object-level hex-labeled supervision. VLM grounding identifies caption-relevant objects, segmentation masks isolate object pixels, and MeanShift clustering in CIELAB space [22] estimates dominant colors under natural illumination. Shadows make these real-image labels approximate colors. To supply a clean low-level color reference, we additionally train on pure-color images as a ‘‘pure-color anchor’’: each sample pairs a hex code with a solid-color target, and these samples are activated only at high-noise timesteps. This anchor gives the model a clean mapping from numeric hex strings to RGB statistics while low-noise training uses natural images. To better evaluate any-color generation and editing, we introduce Any Color Benchmark (ACBench), an object-level color-fidelity benchmark with two components: ACBench-T2I and ACBench-Edit. It measures the color fidelity of generated or edited object regions to the requested hex values. On FLUX.2-4B, Paint-Anything improves ACBench color-fidelity scores by 85.3% for generation and 28.3% for editing relative to the base model, and ablations show that pure-color anchors are an important ingredient in learning reliable pixel-space color control. Beyond ACBench, the same finetuned model achieves the highest average CompColor score [54]

2

Denoising Transformer …

N

Denoising Transformer

t ∈ (0,1)

t ∈ (0.8,1)

t=0 Noise

Noise

VAE Encoder

Text Encoder Wrap hex codes in <color></color>

Text Prompt

Target Image

Target Image

(Pure Color)

(Real World)

Condition Image

Figure 3 Training architecture. Hex colors are represented directly in the text prompt and encoded by the text encoder as text tokens. The VAE encodes target images, and the denoising transformer is trained with the flow-matching objective. Pure-color targets are used only at high-noise timesteps t ∈ [tgate , 1], real-world targets are sampled across t ∈ (0, 1), and optional condition images for editing are encoded as clean image latents at t = 0.

among the compared methods under both named-color and hex prompts. The gains are further supported by GenColorBench NCU evaluation and a human preference study (Appendices A and K). Our contributions are three-fold: • We discover that object-level hex supervision enables unified, prompt-native color control across generation and editing, and strengthen this capability with timestep-gated pure-color grounding. • We develop a data pipeline that converts real images into object-level hex supervision for generation and editing. We use this pipeline to build Paint-500K. • We introduce ACBench to evaluate object-level hex color fidelity in generation and editing, and demonstrate consistent gains across backbones and independent benchmarks.

2

Related Work

2.1

Color control in text-to-image generation

Recent work has begun to make color a first-class control signal in text-to-image generation. Some methods learn or modify color representations, such as learnable color prompts [11], CIELAB-aligned text embeddings [57], or numeric-color encoders for RGB/hex strings [13]. Others use training-free inference-time techniques, such as sampling guidance, attention/value manipulation, or color-alignment objectives [2, 3, 39, 45, 48, 55, 56]. These approaches improve color controllability, but often rely on additional control pathways, expensive inference-time optimization, or task-specific assumptions, which can limit unified task coverage. NumColor [13] learns numerical color embeddings, while BBQ-to-Image [34] uses structured prompts containing bounding boxes and RGB triplets.1 Our model learns hex-to-color grounding through object-level supervision and uses hex codes directly in object descriptions and editing instructions. It unifies generation and editing without a dedicated color encoder or an intermediate layout representation containing bounding-box coordinates.

2.2

Color editing and image colorization

General image-editing systems based on prompt editing, instruction tuning, inversion, in-context editing, flow-based editing, or conditional control [10, 31, 33, 35, 38, 62, 63, 71, 73] can change visual attributes. Here, 1 Public inference code and checkpoints needed to run these methods on ACBench were unavailable at evaluation time. We nevertheless compare with NumColor using its reported GenColorBench NCU results in Appendix A.

3

Blue sky

VLM

cloud

cloud

grounding

Blue sky

CIE Space Cluster

SAM3

house

<Caption>

cloud

cloud

house

Road

Text-to-Image pair

Road

Instances

Multi-object data

Instances Select a dominant-color instance

VLM recaptioning

Single-object data

Blue sky

Edit Expert Color edit

#6998A2 sky

LowLevel Filter

Color Edit data Edited image (as source!)

instance mask

Change sky to #6998A2

Dominant-color instance

Figure 4 Paint-500K data pipeline. Real image-caption pairs are converted into hex-conditioned T2I and editing data through VLM grounding, SAM3 masks, CIE-space color extraction, edit synthesis, filtering, and recaptioning.

we focus on edits specified by numerical color values. More specialized color-editing methods improve languageguided, region-aware, or continuous object color control [26, 58, 65, 66, 68, 69]. Palette- and histogram-based recoloring methods [1, 16, 19] provide explicit controls over image colors, rather than binding hex specifications to objects through text prompts. Image-reference conditioning methods such as IP-Adapter [67] provide continuous visual cues, but global image conditioning can entangle target color with style or texture. In parallel, image colorization methods [4, 5, 17, 18, 23, 40, 41] use masks, palette images, or semantic cues to constrain chromatic output. Our goal is to make the same hex-conditioned prompt interface work across generation and editing with a single finetuned model.

3

Method

3.1

Preliminaries

We build on a rectified-flow text-to-image model [27, 43] with a text encoder, a VAE [36, 52], and a denoising transformer [27]. Let y denote the text prompt and I ⋆ the target image. For editing, I c denotes the source image. The frozen text encoder maps y into embeddings c. The VAE encoder maps the target image into a latent sequence z0 = Evae (I ⋆ ); for editing, the optional source image is encoded by the same VAE into a clean condition latent zc = Evae (I c ). Training follows the flow-matching objective. We sample ϵ ∼ N (0, I) and t ∼ U(0, 1), where t = 0 corresponds to the clean target latent and t = 1 corresponds to pure noise, and form zt = (1 − t)z0 + tϵ,

v ⋆ = ϵ − z0 .

(1)

The denoiser predicts a velocity from the noisy target latent, the timestep, the text tokens, and optional image-condition tokens: bθ = Dθ ([zt ; zc ], t, c) , v (2) where [zt ; zc ] denotes sequence concatenation and reduces to zt for T2I samples. The base loss is h i 2 LFM = Ez0 ,ϵ,t,y ∥b vθ − v⋆ ∥2 .

(3)

Thus generation and editing share the same denoising objective; they differ only in whether the clean condition-image latent is concatenated to the noisy target latent. 4

3.2

Training pipeline

Figure 3 summarizes our training pipeline for grounding a shared hex-prompt interface through object-level supervision and timestep-gated pure-color anchors. The model follows ordinary text prompts for generation and takes a source image for editing. Explicit color-tag wrapping. We mark each 24-bit sRGB hex specification as a color attribute by enclosing

it in <color> and </color> tags directly in the prompt. Because the text encoder remains frozen during finetuning, we make the color span explicit in its input before encoding. For example, prompts can write “a photo of a <color> #CFEFFB </color> colored car” or “change the bag to <color> #FFF8C4 </color>.” Empirically, this wrapping improves object-level color fidelity and compositional color binding, with gains on ACBench-T2I, ACBench-Edit, and CompColor (Table 2). Unified generation and editing format. We jointly train generation and editing so that one model learns color

perception, object-color binding, and color-conditioned editing through the same hex-prompt interface. Every sample contains a target image and a hex-conditioned prompt; editing samples also contain a source image encoded as a clean condition-latent branch. Pure-color anchor supervision. Real images provide object semantics, but their measured colors are noisy:

shadows can map the same object to many plausible RGB values, leaving only an approximate color. We therefore add pure-color anchors: we sample 24-bit RGB values, render each value as a solid-color image, and pair it with a prompt containing the corresponding <color>#HEX</color> token. These anchors provide a clean low-level signal for hex-to-RGB grounding, so we use them to focus training on color-token alignment rather than on object semantics. Motivated by early color stabilization in pure-color generation trajectories (Appendix E), we apply a high-noise gate: pure-color samples are trained only with t ∼ U(tgate , 1.0), while real-image T2I and editing samples use t ∼ U(0, 1). For tgate = 1, pure-color samples are trained at t = 1. Table 3 shows that high-noise gating improves generation and editing over ungated anchors. Training loss. The final training objective combines the three streams:

L = λt2i Lt2i + λedit Ledit + λrgb Lrgb ,

(4)

where each term uses the flow-matching loss in Eq. 3. Lt2i is computed on real-image color-generation samples, Ledit on paired image-editing samples, and Lrgb on pure-color anchor samples under the high-noise gate. We finetune the denoising transformer with the VAE and text encoder frozen. Additional mixture and optimization details are provided in Appendix D.

3.3

Data pipeline Original

K-means K=3

K-means K=5

Extracting object colors. Extracting a representa-

MeanShift h=0.05

tive color from an object is crucial for reliable color supervision: lighting and shadows can make a singlecolored object contain many pixel values. A common approach is RGB-space clustering [34, 47, 66], typically using k-means. This has two limitations: (I) RGB distances do not reflect human perception [22, 66], potentially separating perceptually simFigure 5 MeanShift reduces redundant color splits ilar colors or merging distinct ones; (II) objects have compared with fixed-K k-means. Both methods varying numbers of dominant colors, so a fixed K cluster the outlined skin patch in CIELAB before filtering. can split shading and noise into redundant labels. Maps show cluster membership; bars show mean-RGB Inspired by ColorBind [54], we use the perceptual palettes weighted by pixel share. CIELAB space [22]. We further introduce MeanShift [21] into object-color annotation to adapt the cluster count to each object’s color distribution. Empirically, MeanShift reduces redundant color splits relative to fixed-K k-means in CIELAB, yielding a more coherent dominant-color region (Figure 5). Binding colors to objects and captions. Paint-500K uses high-quality real image-caption pairs from an

internal collection. A VLM identifies caption-relevant objects and their bounding boxes, and SAM3 [15] 5

A two-tone car: top is #FEFEFE, bottom is #FF5634.

A #F4A460 desert with a #000080 oasis.

A #F5FFFA room with a #B22222 door.

A #F0FFEE flower in a snowstorm.

A translucent jelly cube in #ACFF30.

A #AFE5EE sky with #2F4F4F clouds.

Figure 6 Qualitative comparison with the base model. Each pair shows FLUX.2-4B (left) and Paint-Anything (right) on the same hex-conditioned prompt. The baseline often produces colors that are semantically plausible but numerically distant from the requested hex value.

produces the object masks used for color extraction. We filter clusters by pixel coverage and discard instances without sufficient retained coverage. The largest retained cluster provides an sRGB hex label for its object. The VLM then rewrites the caption to bind each label to the corresponding noun phrase using explicit <color>#HEX</color> spans. Figure 4 shows the full pipeline; filtering details and VLM templates are in Appendices D and F. Generation and editing streams. For T2I supervision, multi-object captions teach compositional color binding, while single-object crops emphasize direct object-color perception. We collect 400K T2I samples: 100K single-object and 300K multi-object examples. For editing, a pretrained editing model (Appendix D) recolors each single-object image to form the source, and the original photograph serves as the target, paired with an instruction to restore its hex color. Only the source is generated, so the targets carry no generative artifacts. Filtering for scene preservation, meaningful recoloring, and instruction consistency yields 100K editing samples that use the same prompt-native color syntax.

4

Experiments

4.1

Benchmarks and metrics

CompColor benchmark. CompColor [54] is a compositional color-binding benchmark that measures whether a generator can faithfully bind two distinct colors to two objects in the same prompt. Each prompt has the form “a {color} colored {object} and a {color} colored {object}”, with color words drawn from a fixed named-color palette, e.g., “a paleturquoise colored shirt and a ivory colored bench”. The released baseline table covers a wide range of color-binding methods evaluated under this named-color setting. Our task instead specifies prompt colors as 24-bit hex strings. We therefore evaluate CompColor under an additional hex-translated transfer protocol that replaces each color word with its canonical hex value while keeping the original object compositions intact, e.g., “a #AFEEEE colored shirt and a #FFFFF0 colored bench”. More details about this benchmark and our hex-translated protocol are in Appendix C. Any Color Benchmark (ACBench). To evaluate object-level color fidelity under 24-bit hex prompts in both generation and editing, we introduce ACBench. ACBench-T2I contains 1000 generation prompts over common object categories, and ACBench-Edit contains 500 real-image recoloring prompts. Following prior object-based benchmarks [30], we pair randomly sampled hex colors with everyday objects, animals, and plants. ACBench covers three settings: • Single-object (Single) (500 prompts): one object with one target color, e.g., “a photo of a #CFEFFB car.” 6

ACBench-T2I ↑

ACBench-Edit ↑

CompColor ↑

Total params.

Single

Two

Overall

Edit

Single

Close

Distant

1.0B 17B 10B 27B 17B 56B 8B 8B 10B

15.00 25.00 32.11 26.07 36.24 51.32 37.45 37.45 32.15

8.00 21.00 35.00 24.34 40.07 52.08 36.60 36.60 34.74

11.50 23.00 33.56 25.21 38.15 51.70 37.02 37.02 33.45

— — — 54.40 64.53 68.87 58.90 58.90 —

0.53† 0.56† 0.65 0.59 0.75 0.73 0.74 0.34 —

0.36† 0.54† 0.69 0.59 0.67 0.72 0.70 0.41 —

0.30† 0.49† 0.66 0.63 0.68

Color-specialized methods ColorBind/Edit CtrlColor ColorPeel ColorWave*

var. SD1.5-based SD1.4-based SDXL-based

30.42 — 45.28 50.36

34.17 — 34.63 42.71

32.30 — 39.96 46.54

60.38 57.46 — —

0.72 — 0.68 0.72

0.71 — 0.62 0.68

0.73 — 0.64 0.70

FLUX.2-4B + Ours FLUX.2-4B + Ours (Hex) Z-Image Base + Ours

8B 8B 10B

72.67 72.67

64.49 64.49

68.58 68.58

75.57 75.57

0.81

0.80 0.80

0.79

56.76

50.78

53.77

—

0.77

0.76

Model SD1.5 FLUX.1 Z-Image-Turbo Qwen-Image / Edit FLUX.2-9B FLUX.2-dev FLUX.2-4B FLUX.2-4B (Hex) Z-Image Base

0.78 0.77

0.79 0.73 0.38 —

0.76

Table 1 Main results: ACBench (0–100) and CompColor (0–1). CompColor uses named colors unless marked (Hex); paired rows share ACBench results. —: unreported. †: quoted from ColorBind [54]. * : our reproduction. Reruns use official defaults (FLUX.2-4B Edit CFG 2.0). NCU and additional CompColor results: Appendices A and C.

• Two-object (Two) (500 prompts): two objects, each assigned a distinct target color, e.g., “a photo of a #FFF8C4 dog and a #CFEFFB chair.” • Edit (500 prompts): one object with one target color, e.g., “change the bag to #CFEFFB .” Protocol and metric. We use SAM3 [15] to obtain the target masks. For T2I, we segment the prompted object in the generated image. For editing, we segment the target object in the source image, inspect the mask manually, and reuse the same source mask across the compared edited outputs. For a target P mask M , let c̄ = |M |−1 p∈M I(p) be the mean sRGB vector on the 0 to 255 scale. Following prior work [12, 54], estimated object color with the target. Given target RGB c⋆ , we compute P we compare the 1 ⋆ MAE = 3 k∈{R,G,B} |c̄k − ck |; failed localization receives a score of 0. CIELAB estimators and CIEDE2000 evaluation appear in Appendices H and A. We convert MAE into a normalized score using a piecewise linear function:   max(0, MAE − 16) . s = 100 · max 0, 1 − 48

(5)

The score measures region-level color fidelity while allowing natural appearance variation. MAE of at most 16 receives full score; this threshold applies to the average absolute channel error. MAE between 16 and 64 is linearly penalized, and MAE of at least 64 receives 0. The linear interval retains graded credit for imperfect color matches, preserving distinctions that a tighter cutoff would collapse to zero. We generate one image per T2I prompt using one seed per prompt. We report separate scores for the equally sized Single and Two splits, with Overall as their arithmetic mean. We also conduct a user study to examine whether people prefer our color results to those of the base model (Appendix K).

4.2

Experimental setup

Baseline and implementation details. We finetune FLUX.2-klein-base-4B [8] and Z-Image Base [70] with

the same training recipe. Both models are trained for 4000 steps on 4 GPUs, using Adam with a global batch size of 72 and a learning rate of 2 × 10−5 . Throughout the paper, FLUX.2-4B and FLUX.2-9B denote the 7

#5B262F

#873250

#AD3351

#D03F47

#F34B41

#FF685E

#5AF5E5

#6FF2CE

#A7F6C8

#ABF8B1

#ADF9A3

#A3FF66

Figure 7 Hex-conditioned generation. Paint-Anything follows a range of requested hex colors for the same object. All results use the same random seed.

Original Image

#FFB2B9

#FFC3C5

#FDBCDB

#FCB1F5

#ED94FC

Original Image

#FB8214

#FF7B32

#FF9B00

#FEB000

#FFCD00

Figure 8 Hex-conditioned editing. Given a source image and a hex-color instruction, Paint-Anything follows the requested hex color when recoloring the target object.

undistilled klein base checkpoints.2 The FLUX ablations use the 4B backbone with the same optimization settings. Table 1 compares complete systems; same-backbone ablations assess our training recipe. The open-source baselines are SD1.5 [52], FLUX.1 [6], Z-Image-Turbo [70], Qwen-Image and QwenImage-Edit [60], FLUX.2-4B [8], FLUX.2-9B [9] and FLUX.2-dev [7]; the color-specialized baselines are ColorBind/Edit [54], CtrlColor [41], ColorPeel [11] and ColorWave [39]. NumColor inference code and checkpoints were unavailable at evaluation time, so we use its published NCU results (Appendix A). CompColor includes named-color and hex prompts (Appendix C). Evaluation details.

4.3

Quantitative results

Table 1 shows that our finetuned FLUX.2-4B improves over its base checkpoint across three complementary aspects of color control: object-level hex color fidelity for text-to-image generation on ACBench-T2I, recoloring fidelity for image editing on ACBench-Edit, and compositional color binding on CompColor. Hex supervision enables an 8B model to outperform a 56B model. Within the FLUX.2 family, ACBench-T2I 2 The “4B” and “9B” suffixes refer to the DiT parameter count; the Total params. column in Table 1 reports the total parameter count, including the bundled Qwen3 text encoder.

8

Overall and ACBench-Edit scores increase across the off-the-shelf models with 8B, 17B, and 56B total parameters. Yet our finetuned 8B model exceeds the 56B model by 16.88 points on ACBench-T2I and 6.70 points on ACBench-Edit. The same training recipe also produces a large gain on Z-Image Base. Finetuning closes the word-to-hex gap. On CompColor, replacing color names with raw hex strings reduces

the FLUX.2-4B base model’s average score from 0.72 to 0.38. After localized hex supervision, the hex-prompt average more than doubles and exceeds the base model’s named-color average. The named-color average also improves from 0.72 to 0.79, showing that finetuning preserves and improves the model’s existing compositional color-binding ability. Paint-Anything outperforms the evaluated specialized systems. On ACBench, Paint-Anything exceeds the

strongest specialized baseline in each task: ColorWave by 22.04 points on ACBench-T2I and ColorBind/Edit by 15.19 points on ACBench-Edit. The independent GenColorBench NCU evaluation also ranks Paint-Anything highest among the compared methods, while a human preference study favors its color results over those of the base model (Appendices A and K).

4.4

Ablation study

We ablate four choices in the final recipe: color-token wrapping, pure-color anchors, highnoise gating, and full-model finetuning. Table 2 shows that the complete recipe is best across ACBench generation, ACBench editing, and CompColor, while Table 3 isolates the gate threshold. Full finetuning and LoRA. With

Base

LoRA

Wrap

Pure

Gate

T2I ↑

Edit ↑

CompC. ↑

✓

—

—

—

—

37.02

58.90

0.38

× × × × × ×

× × × × ✓ ×

× ✓ ✓ × ✓ ✓

× ✓ × ✓ ✓ ✓

× × × ✓ ✓ ✓

44.06 63.83 57.16 48.01 48.45

65.81 70.39 73.89 64.08 58.71

0.52 0.71 0.71 0.59 0.54

68.58

75.57

0.79

Table 2 Module ablations on FLUX.2-4B. Base denotes the pretrained model without finetuning; otherwise, LoRA × denotes full finetuning and ✓ denotes rank-256 LoRA. Pure denotes pure-color anchors. T2I/Edit use ACBench; CompC. is the hex-prompt CompColor average.

the complete recipe, a shared learning rate and 4000 steps, full finetuning exceeds rank-256 LoRA: 68.58 vs. 48.45 (T2I), 75.57 vs. 58.71 (Edit), and 0.79 vs. 0.54 (CompColor).

Wrapping clarifies hex syntax, while anchors provide a clean color reference. Without pure-color anchors,

bare-hex finetuning reaches only 44.06 and 65.81 on ACBench-T2I and ACBench-Edit. Color-token wrapping raises these scores to 57.16 and 73.89, giving the model a clearer textual handle for raw hex strings. With wrapping enabled, adding pure-color anchors provides a clean hex-to-RGB reference; without gating, generation improves to 63.83 but editing decreases to 70.39. The ungated anchor objective therefore helps generation at the cost of editing fidelity. tgate

0.0

0.7

0.8

0.9

68.58 75.57 0.79

67.54 73.68 0.77

1.0

High-noise gating makes anchor supervision more effective. Restricting pure-

color supervision to high-noise timesteps matches where color emerges in pure0.79 color trajectories (Appendix E) and leaves lower-noise training to real-image generaTable 3 Timestep-gate ablation on FLUX.2-4B, with wrapping and tion and editing. With wrapping enabled, anchors enabled. tgate = 0 matches the ungated anchor row in Table 2. our gated recipe improves ACBench-T2I by 4.75 points and ACBench-Edit by 5.18 points over ungated anchors, while also improving CompColor. It also outperforms training without anchors on both generation and editing. Thresholds 0.7, 0.8, and 0.9 all improve both tasks over ungated anchors (Table 3), showing low sensitivity. We use tgate = 0.8. ACBench-T2I ↑ ACBench-Edit ↑ CompColor ↑

63.83 70.39 0.71

65.93 72.68

67.62 71.49 0.76

9

4.5

Qualitative results

Paint-Anything better matches the requested hex values across object categories and color families, whereas FLUX.2-4B often produces plausible but numerically inaccurate colors under the same prompts (Figure 6). Figure 7 shows generation for the same object under different hex specifications, while Figure 8 shows recoloring a target object in a source image according to a hex-color instruction.

5

Limitations and Future Work

Our current training data do not include palette-specific supervision. Future work can add palette-specific supervision and extend the same unified interface to a broader range of color-control tasks.

6

Conclusion

We presented Paint-Anything, showing that pretrained image models can learn any-color control as a unified prompt-native capability for generation and editing. Paint-500K supplies object-level hex supervision through perceptual color clustering, with our empirical analysis motivating MeanShift over fixed-K clustering. Timestep-gated pure-color anchors add 4.75 and 5.18 points on ACBench-T2I and ACBench-Edit over ungated anchors. Overall, Paint-Anything improves ACBench color fidelity by 85.3% for generation and 28.3% for editing relative to the base model. On the independent CompColor and GenColorBench NCU benchmarks, Paint-Anything achieves the highest average scores among the compared methods, supporting its compositional color binding and numerical color understanding.

10

References [1] Mahmoud Afifi, Marcus A. Brubaker, and Michael S. Brown. Histogan: Controlling colors of gan-generated and real images via color histograms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7941–7950, 2021. 4 [2] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Training-free color-style disentanglement for constrained text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 6291–6300, 2025. 3 [3] Elad Aharoni, Noy Porat, Dani Lischinski, and Ariel Shamir. Palette aligned image diffusion. Computer Graphics Forum, page e70384, 2026. 1, 3 [4] Yanru An, Ling Gui, Qiang Hu, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, and Yanfeng Wang. Controllable image colorization with instance-aware texts and masks, 2025. 4 [5] Hyojin Bahng, Seungjoo Yoo, Wonwoong Cho, David K. Park, Ziming Wu, Xiaojuan Ma, and Jaegul Choo. Coloring with words: Guiding image colorization through text-based palette generation. In Proceedings of the European Conference on Computer Vision, 2018. 4 [6] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 8 [7] Black Forest Labs. FLUX.2 [dev] model card. https://huggingface.co/black-forest-labs/FLUX.2-dev, 2025. Accessed: 2026-04-16. 8 [8] Black Forest Labs. FLUX.2 [klein] Base 4B model card. https://huggingface.co/black-forest-labs/FLUX. 2-klein-base-4B, 2026. Accessed: 2026-04-16. 7, 8 [9] Black Forest Labs. FLUX.2 [klein] Base 9B model card. https://huggingface.co/black-forest-labs/FLUX. 2-klein-base-9B, 2026. Accessed: 2026-04-16. 8 [10] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023. 3 [11] Muhammad Atif Butt, Kai Wang, Javier Vazquez-Corral, and Joost van de Weijer. Colorpeel: Color prompt learning with diffusion models via color and shape disentanglement. In Proceedings of the European Conference on Computer Vision, 2024. 1, 3, 8 [12] Muhammad Atif Butt, Alexandra Gomez-Villa, Tao Wu, Javier Vazquez-Corral, Joost van de Weijer, and Kai Wang. GenColorBench: A color evaluation benchmark for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 36638–36648, 2026. 7, 15, 16, 21 [13] Muhammad Atif Butt, Diego Hernández, Alexandra Gomez-Villa, Kai Wang, Javier Vazquez-Corral, and Joost Van De Weijer. Numcolor: Precise numeric color control in text-to-image generation. In European Conference on Computer Vision, 2026. Accepted for publication. 3, 15 [14] ByteDance Seed Team. Seed1.8 model card. https://github.com/ByteDance-Seed/Seed-1.8, 2026. Accessed: 2026-05-04. 18 [15] Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. SAM 3: Segment anything with concepts. In International Conference on Learning Representations, 2026. 5, 7, 21 [16] Huiwen Chang, Ohad Fried, Yiming Liu, Stephen DiVerdi, and Adam Finkelstein. Palette-based photo recoloring. ACM Transactions on Graphics, 34(4):139:1–139:11, 2015. 4 [17] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-cad: Language-based colorization with any-level descriptions using diffusion priors. In Advances in Neural Information Processing Systems, pages 77174–77186, 2023. 4 [18] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-coins: Language-based colorization with instance awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19221–19230, 2023. 4

11

[19] Cheng-Kang Ted Chao, Jason Klein, Jianchao Tan, Jose Echevarria, and Yotam Gingold. Colorfulcurves: Paletteaware lightness control and color editing via sparse optimization. ACM Transactions on Graphics (TOG), 42(4): 1–12, 2023. 4 [20] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics, 2023. 18 [21] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603–619, 2002. 5 [22] Commission Internationale de l’Eclairage. Colorimetry, 4th edition. Technical Report CIE 015:2018, CIE, 2018. 2, 5 [23] Xiaoyan Cong, Yue Wu, Qifeng Chen, and Chenyang Lei. Automatic controllable colorization via imagination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2609–2619, 2024. 4 [24] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. In European Conference on Computer Vision, 2024. 18 [25] Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025. 1 [26] Yi Dong, Yuxi Wang, Ruoxi Fan, Wenqi Ouyang, Zhiqi Shen, Peiran Ren, and Xuansong Xie. Chromafusionnet (cfnet): Natural fusion of fine-grained color editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1591–1599, 2024. 4 [27] 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. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the International Conference on Machine Learning, 2024. 1, 4 [28] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In International Conference on Learning Representations, 2023. 18 [29] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545–7556, 2023. 18 [30] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132–52152, 2023. 6, 21 [31] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In International Conference on Learning Representations, 2023. 3 [32] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation. In Advances in Neural Information Processing Systems, 2023. 21 [33] Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang, and Renjie Liao. Uniedit-flow: Unleashing inversion and editing in the era of flow models. In International Conference on Learning Representations, 2026. 3 [34] Eliran Kachlon, Alexander Visheratin, Nimrod Sarid, Tal Hacham, Eyal Gutflaish, Saar Huberman, Hezi Zisman, David Ruppin, and Ron Mokady. Bbq-to-image: Numeric bounding box and qolor control in large-scale text-toimage models, 2026. 3, 5 [35] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007–6017, 2023. 3 [36] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, 2014. 4 [37] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023. 17

12

[38] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19721–19730, 2025. 3 [39] Héctor Laria, Alexandra Gomez-Villa, Jiang Qin, Muhammad Atif Butt, Bogdan Raducanu, Javier VazquezCorral, Joost van de Weijer, and Kai Wang. Leveraging semantic attribute binding for free-lunch color control in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7689–7698, 2026. 1, 3, 8 [40] Yifan Li, Yuhang Bai, Shuai Yang, and Jiaying Liu. Coco-lc: Colorfulness controllable language-based colorization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10939–10947, 2024. 4 [41] Zhexin Liang, Zhaochen Li, Shangchen Zhou, Chongyi Li, and Chen Change Loy. Control color: Multimodal diffusion-based interactive image colorization, 2024. 4, 8 Qwen-Image-Edit-2511-Lightning model card. [42] LightX2V Team. Qwen-Image-Edit-2511-Lightning, 2026. Accessed: 2026-05-04. 18

https://huggingface.co/lightx2v/

[43] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In International Conference on Learning Representations, 2023. 4 [44] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In European Conference on Computer Vision, 2024. 17 [45] Alexander Lobashev, Maria Larchenko, and Dmitry Guskov. Color conditional generation with sliced wasserstein guidance. In Advances in Neural Information Processing Systems, 2025. 1, 3 [46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations, 2024. 1 [47] Pylette Developers. Pylette: A python library for extracting color palettes from images. https://pypi.org/ project/pylette/, 2026. Accessed: 2026-05-05. 5 [48] Qianru Qiu, Jiafeng Mao, and Xueting Wang. Exploring palette based color guidance in diffusion models. In Proceedings of the 33rd ACM International Conference on Multimedia, 2025. 1, 3 [49] Qwen Team. Qwen-Image-Edit-2511 model card. https://huggingface.co/Qwen/Qwen-Image-Edit-2511, 2025. Accessed: 2026-05-04. 18 [50] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. In Advances in Neural Information Processing Systems, pages 3536–3559, 2023. 18 [51] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded SAM: Assembling open-world models for diverse visual tasks, 2024. 21 [52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 4, 8 [53] Chenxi Ruan, Yihan Hou, Yu Xiao, Guosheng Hu, and Wei Zeng. Colorconceptbench: A benchmark for probabilistic color-concept understanding in text-to-image models, 2026. 15 [54] Shay Shomer-Chai, Wenxuan Peng, Bharath Hariharan, and Hadar Averbuch-Elor. Color bind: Exploring color perception in text-to-image models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1916–1925, 2026. 2, 5, 6, 7, 8, 15, 16, 17, 18, 21 [55] Tripti Shukla, Srikrishna Karanam, and Balaji Vasan Srinivasan. Test-time conditional text-to-image synthesis using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4559–4568, 2026. 1, 3 [56] Ka Chun Shum, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Color alignment in diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1, 3 [57] Sung-Lin Tsai, Bo-Lun Huang, Yu-Ting Shen, Cheng Yu Yeo, Chiang Tseng, Bo-Kai Ruan, Wen-Sheng Lien, and Hong-Han Shuai. Color me correctly: Bridging perceptual color spaces and text embeddings for improved diffusion

13

generation. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 10074–10082, 2025. 3 [58] Zhenwei Wang, Nanxuan Zhao, Gerhard Hancke, and Rynson W. H. Lau. Language-based photo color adjustment for graphic designs. ACM Transactions on Graphics, 42(4):101:1–101:16, 2023. 4 [59] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. TokenCompose: Text-to-image diffusion with token-level supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 21 [60] Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025. 1, 8, 18 [61] Ji Xie, Luke Zettlemoyer, Xudong Wang, et al. Reconstruction alignment improves unified multimodal models. In International Conference on Learning Representations, pages 120095–120137, 2026. 1 [62] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 3 [63] Sihan Xu, Ziqiao Ma, Yidong Huang, Honglak Lee, and Joyce Chai. Cyclenet: Rethinking cycle consistency in textguided diffusion for image manipulation. Advances in Neural Information Processing Systems, 36:10359–10384, 2023. 3 [64] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, et al. Qwen3 technical report, 2025. 2 [65] Yuqi Yang, Dongliang Chang, Yuanchen Fang, Yi-Zhe Song, Zhanyu Ma, and Jun Guo. Controllable-continuous color editing in diffusion model via color mapping, 2025. 4 [66] Yuqi Yang, Dongliang Chang, Yijia Ling, Ruoyi Du, and Zhanyu Ma. Recolour what matters: Region-aware colour editing via token-level diffusion. In European Conference on Computer Vision, 2026. Accepted for publication. 4, 5 [67] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4 [68] Xingxi Yin, Zhi Li, Jingfeng Zhang, Chenglin Li, and Yin Zhang. Coloredit: Training-free image-guided color editing with diffusion model, 2024. 4 [69] Zixin Yin, Xili Dai, Ling-Hao Chen, Deyu Zhou, Jianan Wang, Duomin Wang, Gang Yu, Lionel M. Ni, Lei Zhang, and Heung-Yeung Shum. Training-free text-guided color editing with multi-modal diffusion transformer. In International Conference on Learning Representations, 2026. 1, 4 [70] Z-Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Aiming Hao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Yuming Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Qilong Wu, Feng Yu, Zechao Zhan, Chi Zhang, Shifeng Zhang, Ruikai Zhou, and Shilin Zhou. Z-Image: An efficient image generation foundation model with single-stream diffusion transformer, 2025. 7, 8 [71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 3 [72] Yabo Zhang, Kunchang Li, Dewei Zhou, Xinyu Huang, and Xun Wang. Images in sentences: Scaling interleaved instructions for unified visual generation. arXiv preprint arXiv:2605.12305, 2026. 1 [73] Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. Enabling instructional image editing with in-context generation in large scale diffusion transformer. In Advances in Neural Information Processing Systems, pages 154372–154404, 2025. 3 [74] Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. arXiv preprint arXiv:2410.12669, 2024. 1 [75] Dewei Zhou, Xinyu Huang, Xun Wang, Ji Xie, Yabo Zhang, Liang Li, Kunchang Li, Zongxin Yang, and Yi Yang. Metapoint: Unlocking precise spatial control in agentic visual generation. In European Conference on Computer Vision, pages 373–390. Springer, 2026. 1

14

Supplementary Material Overview We provide additional information in the supplementary material, as outlined below: • Sec. A: How ACBench relates to existing color benchmarks, with the GenColorBench NCU transfer check. • Sec. B: ACBench construction, prompt sampling and baseline settings. • Sec. C: The hex-translated CompColor protocol and the full original baseline roster. • Sec. D: Paint-500K composition, the training mixture and the tools used to build it. • Sec. E: Pure-color generation trajectories motivating the high-noise gate. • Sec. F: Prompt templates for grounding, filtering and recaptioning. • Sec. G: Evaluator settings used for every reported score. • Sec. H: ACBench reliability under alternative region statistics, SAM3 thresholds and localization failures. • Sec. I: Pure-color anchor resolution ablation. • Sec. J: Editing performance as a function of source-to-target color distance. • Sec. K: Human preference study protocol and results. • Sec. L: Training and evaluation separation check. • Sec. M: Additional training-data examples.

A

Relation to Existing Color Benchmarks

Existing benchmarks expose related weaknesses in color-controllable generation. ColorBind introduces CompColor for compositional color binding [54], GenColorBench broadens evaluation to many color-control settings including numerical colors [12], and ColorConceptBench targets implicit color-concept understanding [53]. ACBench complements these efforts by focusing on explicit 24-bit hex-to-object fidelity for both generation and editing. GenColorBench NCU. To check that ACBench gains are not an artifact of our mean-RGB protocol, we

evaluate under the official GenColorBench Numerical Color Understanding (NCU) protocol [12]: four images per RGB/HEX prompt at each model’s default sampling steps and resolution, GroundingDINO+SAM2 localization, OneHue dominant-color extraction, and CIEDE2000 scoring with the released color-neighborhood tolerance. Table 4 reports the result alongside the per-category scores published by NumColor [13], whose evaluation setting we follow. Model FLUX.2-4B Base Paint-Anything FLUX + NumColor† SD3.5 + NumColor† SD3 + NumColor† PixArt-Σ + NumColor† PixArt-α + NumColor†

L1

L3

CSS3/X11

Avg. ↑

34.71 58.15 55.71 51.72 49.23 48.56 44.93

34.64 58.54 48.02 43.35 39.46 40.83 36.17

32.64 56.98 51.96 46.87 47.19 43.90 38.51

34.00 57.89 51.90 47.31 45.29 44.43 39.87

Table 4 GenColorBench NCU, following the evaluation setting of NumColor [13]. L1/L3 denote ISCC-NBS levels; Avg. is their arithmetic mean with CSS3/X11. †: scores quoted from NumColor [13], Table 1, not rerun here. All rows use the same three categories, so the averages are directly comparable.

On the same FLUX.2-4B backbone, finetuning improves mean NCU from 34.00 to 57.89, a gain of 23.89 points, with improvements in all three categories. Our mean also exceeds the published FLUX + NumColor result (51.90) by 5.99 points. GenColorBench NCU remains a single-object T2I numerical-color test; ACBench also covers two-object binding and image editing, while CompColor provides an independent compositional-binding diagnostic. 15

B

ACBench Construction and Evaluation

Task composition and prompts. ACBench-T2I contains 1000 prompts over common object categories: 500

specify one object and one target hex color, and 500 specify two objects with distinct target hex colors. ACBench-Edit contains 500 real-image recoloring prompts, each specifying one target object and its requested hex color. The prompt forms are illustrated in Section 4. Evaluation prompts are constructed independently of Paint-500K captions and use different formats; the exact-string overlap check is reported in Appendix L. Editing images and masks. The 500 editing source images come from an internal dataset of high-quality real

images that does not overlap with the Paint-500K source images. Masking follows Section 4; reusing one manually inspected source mask holds the evaluated region fixed across methods.

Color targets and score interpretation. Target colors are sampled at random from the 24-bit RGB space. Each

target hex is attached to its corresponding object; two-object prompts contain distinct target colors. Scoring uses the MAE-to-score mapping of Section 4; the estimator and localization sensitivity analyses appear in Appendix H. Relation to prior evaluation protocols. Following the object-centric evaluation approach of CompColor [54]

and GenColorBench [12], we localize the requested object before assessing its color. ACBench uses its own SAM3 and sRGB-MAE protocol. For independent evaluation, we use the released CompColor evaluator without modification and the official GenColorBench NCU protocol, including its dominant-color extraction and CIEDE2000-based scoring (Appendices C and A). Thus, the reported gains are evaluated under both our region-mean score and established color-benchmark protocols.

B.1

Baseline Settings and Result Sources

For reproduced baselines, we follow the default configurations of their official open-source implementations and Hugging Face releases, except for the matched FLUX.2-4B editing CFG of 2.0 stated in the main paper. Prompt interfaces and result provenance. ACBench specifies numerical target colors, while CompColor uses

named-color prompts unless a row is marked “Hex.” The latter replaces the color names with their canonical hex values while preserving object compositions. Color-specialized methods retain their respective model backbones and control interfaces; ColorWave is our reproduction, as marked in Table 1. The SD1.5 and FLUX.1 CompColor scores in Table 1 are quoted from ColorBind, as marked on those cells; their ACBench scores are separate evaluations. Additional quoted CompColor baselines appear in Appendix C. NumColor NCU results are arithmetic means of its three published category scores, rather than ACBench reruns (Appendix A). Cross-model results compare complete systems; same-backbone finetuning and ablation comparisons assess our training recipe.

C

CompColor Hex Protocol

Benchmark background. CompColor is the compositional color-binding benchmark released by ColorBind [54],

originally designed to expose a long-standing failure mode of T2I models: when a prompt mentions multiple colored objects, generators often leak color across objects, swap the two assignments, or collapse both objects to a shared dominant hue. CompColor isolates this failure mode by holding the textual structure fixed, varying only the (color, object) bindings, and measuring per-object color fidelity. Each prompt follows the template “a {color1} colored {object1} and a {color2} colored {object2}”. This fixed structure reduces linguistic variation and makes the benchmark a focused test of object–color binding.

Color palette and pair construction. The released benchmark draws color words from a fixed named-color

palette (e.g., tomato, royal-blue, light-cyan, hot-pink) whose canonical RGB triplets are known. For each pair of color names, the perceptual distance between their canonical sRGB values is computed in CIELAB ∗ space (denoted ∆Eab in the original paper). Color pairs are then partitioned by this distance into a Close split, where the two colors are perceptually similar and easy to confuse (e.g., sky-blue vs. light-cyan), and a Distant split, where the two colors are perceptually well-separated (e.g., sky-blue vs. hot-pink). This split is the central design choice of CompColor: Close stresses fine-grained discrimination between adjacent regions of color space, while Distant stresses correct assignment under high color contrast. The benchmark 16

also includes two Single subsets (one color × one object), used to verify that single-object color rendering still works in isolation. Object compositions. Object pairs are sampled from a curated list of common everyday categories (animals, vehicles, garments, household items, food) so that the requested colors are physically plausible and segmentable. The original benchmark fixes both color words and object words up front and ships the prompts as a closed set, so all baselines are evaluated on exactly the same prompts and seeds. We respect this convention: our hex protocol modifies only the color tokens, never the object tokens or pair structure. Evaluation metric. The original CompColor metric estimates, for each generated image, whether each requested

object was rendered in the requested color. In practice this combines an object localizer (a segmenter applied to each object phrase) with a color-similarity check between the segmented region and the target RGB value. The released scores are normalized to a 0 to 1 scale, where higher means better object-level color binding. For CompColor results that we evaluate, including our named-color and hex-translated rows, we run the ColorBind open-source evaluation codebase [54] without modification. Concretely, the pipeline uses LangSAM, which combines Grounding DINO [44] with SAM [37], to locate each object from its text label and produce a segmentation mask; color scoring then follows the original k-means-plus-∆ECMC protocol. This ensures a consistent evaluation protocol across all models. The SD1.5 and FLUX.1 named-color scores in Table 1 are taken from the original release and marked with †; additional quoted baselines are listed in Table 6. Hex translation. We keep the original object compositions and the Single/Close/Distant splits untouched, and

replace only the color words in each prompt with explicit 24-bit hex strings derived from the canonical RGB value of the original color name (e.g., tomato → #FF6347, royal-blue → #4169E1). Named-color baselines and our hex-conditioned rows therefore describe the same perceptual targets, so within-model differences isolate the change from color words to hex specifications. This is a stricter test than CompColor’s named-color setting: hex strings remove the explicit named-color cue (“tomato” suggests a reddish hue), while retaining any numerical-color knowledge already present in the pretrained text encoder. Subset aggregation. The released benchmark provides two single-object subsets evaluated separately in the

original paper. To match the score format used by the most recent baseline tables, we aggregate them by averaging into a single Single column in Table 1; the Close and Distant columns are reported as is. In Table 6 we further report the unweighted mean of Single, Close, and Distant as an Avg column, matching the original paper’s averaging convention. Three-object extension. CompColor was originally proposed as a two-object compositional benchmark, and

the released baseline scores are reported only for this two-object setting. We therefore keep the main-paper table aligned with those released two-object baselines. As a complementary diagnostic, we further evaluate the same wrapped-hex interface on CompColor’s three-object extension (Table 5). Protocol Valid disjoint masks Missing/overlapping masks score zero

Paint-Anything

FLUX.2-4B Base

0.763 0.506

0.362 0.210

Table 5 CompColor three-object extension evaluated under the wrapped-hex interface.

Paint-Anything remains ahead under both protocols. The evaluator scores the designated target object, so this experiment measures color control in scenes containing three colored objects rather than strict three-way joint correctness. Score format. All CompColor hex-protocol scores follow the original 0 to 1 convention, while ACBench scores

follow our 0 to 100 percentage-point scale. The two scales are intentionally kept distinct in Table 1 (with explicit ↑ markers) so that no direct numerical comparison across benchmarks is implied; CompColor measures compositional binding on a closed prompt set, while ACBench measures object-level hex color fidelity on broader generation and editing protocols. Full original baseline roster. The main paper Table 1 keeps the modern T2I/edit baselines that overlap with ACBench so that the same backbone family can be tracked across both benchmarks. For completeness, Table 6

17

reports the remaining original CompColor baselines released by ColorBind [54], namely Attend-and-Excite [20], Structured Diffusion [28], SynGen [50], RichText [29] and Bounded Attention [24], with the same (Single, Close, Distant, Avg) format as the original paper. These rows are copied directly from the release; they are included as a reference for the named-color regime and are not re-evaluated under our hex-translated protocol. We do not assess whether each method could be adapted to hex inputs. Method

Single ↑

Close ↑

Distant ↑

Avg ↑

0.51 0.61 0.55 0.58 0.45 0.29 0.43

0.38 0.33 0.46 0.39 0.49 0.35 0.55

0.28 0.26 0.38 0.39 0.56 0.30 0.29

0.39 0.40 0.46 0.45 0.50 0.31 0.42

SD 1.4 SD 2.1 Attend-and-Excite Structured Diffusion SynGen RichText Bounded-Attention

Table 6 Remaining original CompColor baselines from ColorBind [54], omitted from the main paper Table 1. Scores follow the released 0 to 1 convention.

D

Training Data and Mixture

We set λt2i = λedit = λrgb = 1 in all experiments. Paint-500K contains 500K color-control samples built from an internal collection of high-quality real images. The source collection is not publicly available. It is organized into two streams: 400K T2I generation samples and 100K editing samples. The T2I stream contains 100K single-object color generation samples and 300K multi-object color generation samples. Rows include target image paths, object prompts, source image paths for editing, object masks, target hex colors, and dominant-color metadata. We use Seed1.8 [14] as the VLM for grounding, filtering, and recaptioning, and use Qwen-Image-Edit-2511 [49, 60] with the 4-step Qwen-Image-Edit-2511-Lightning LoRA [42] as the editing model. Final training uses three streams in total: Paint-500K T2I generation, Paint-500K editing, and auxiliary pure-color grounding. Each training batch contains 30 real-image T2I samples, 12 pure-color samples, and 30 editing samples, giving a total batch size of 72. The auxiliary pure-color stream contains 10K solid-color samples at 512 × 512 resolution and is sampled only for timesteps t ∈ [0.8, 1.0] in the final recipe. All 4B finetuning and ablation rows use Adam, learning rate 2 × 10−5 , and 4000 steps. For FLUX.2-klein-base-4B, T2I generation uses the default classifier-free guidance of 4.0 unless otherwise noted. All FLUX.2-4B editing comparisons, including the base model and finetuned variants, use the same CFG of 2.0. Construction details. For instance grounding, each VLM proposal contains a text label, a short region

description, and a bounding box; SAM3 then segments each grounded box so that color extraction is performed on object pixels. We normalize the CIELAB channels to [0, 1] and cluster masked pixels with MeanShift using bandwidth 0.05. Two filters then decide which colors are retained: each retained cluster must cover at least 15% of the object pixels (COLOR_MIN_RATIO = 0.15), and the retained clusters must jointly cover at least 90% of the object pixels (INSTANCE_KEEP_RATIO = 0.90); otherwise the instance is discarded. Small highlight, shadow, or decoration regions are therefore removed, while multiple substantial colors on one object can be retained. When a single dominant color is required, we take the retained cluster with the largest pixel mass and convert it to a 24-bit sRGB hex label. As a spot check of dominant-color labels, we randomly inspected 100 Paint-500K samples; 96 labels agreed with the human-perceived dominant object color. When several objects share the same semantic label, VLM grounding returns separate bounding boxes for each instance, including same-category objects with different colors (Figure 4). Color extraction is performed independently inside each box. During recaptioning, the localized instances, extracted colors, and spatial relations are aggregated into a complete caption, enabling descriptions such as a red object on the left and 18

a green object of the same category on the right. Multi-object T2I prompts are produced by feeding the original caption, grounded layout, text labels, and extracted hex colors back to the VLM, which inserts exact <color>#HEX</color> spans into the relevant noun phrases. Single-object T2I samples are constructed from instances whose dominant color covers a large fraction of the mask: we randomly expand the instance box, crop a local image region, and recaption the crop around the selected object and its dominant hex color. For editing data, we keep a synthesized source-target pair only when the mask-outside change is small enough to preserve the surrounding scene and the mask-inside change is large enough to verify a meaningful color edit. VLM verification and recaptioning further require that the target object remains identifiable, the requested color change is visually grounded, and any visible texture, material, pattern, or shape changes are reflected in the final instruction while preserving the exact target hex string.

E

Pure-Color Generation Trajectories 1.0

0.95

0.90

0.85

0.80

0.75

0.70

0.50

0.25

0.0

xt

x0

xt

x0

Figure 9 Early color stabilization in pure-color generation. Two trajectories visualize noisy states (rows labeled xt ) and corresponding single-step denoised predictions (rows labeled x0 ) at decreasing noise levels. In these examples, the predicted color is already visually stable near t ≈ 0.8. These pure-color trajectories motivate concentrating anchor supervision at high noise; Table 3 evaluates the resulting choice on generation and editing.

Figure 9 explores when color emerges in two pure-color generation trajectories. It provides a qualitative motivation for high-noise anchor supervision. The downstream evidence comes from the gate ablation in Table 3, where tgate = 0.8 improves both generation and editing over ungated anchors.

F

Prompt Templates and Data-Generation Tools

We used fixed prompt templates for VLM-based caption refinement, instance grounding, and edit-instruction refinement. The color-caption refinement system prompt was: Role: You are a visual data refinement expert specializing in integrating precise color attributes into image captions for VLM training. Task: Integrate detected objects and their specific HEX color codes into a global caption. Use the provided bounding box coordinates only to determine spatial relationships and distinguish between multiple objects, but do not include the numerical coordinates in the final output. Constraints & Rules: No Bbox in Output: Do not include any coordinate values (x1, y1, x2, y2) or bbox IDs in the final caption. Use them only to infer positions

19

(e.g., "on the left," "in the background"). Preserve Context: Maintain the original global caption's narrative flow and structure. Disambiguation: If multiple similar objects exist, use their relative spatial positioning to make the description unique (e.g., "The <color>#HEX</color> bottle on the right"). Natural Integration: Insert the color tags directly before or after the object they describe so the sentence remains grammatically fluent. Return Format: "[Final enhanced caption text with <color>HEX code</color> tags and no coordinates]" Strict HEX Usage: You must use the EXACT HEX color codes provided in the input. Do not invent or hallucinate colors that are not in the input (e.g., #000000, #FFFFFF). Example for your reference: User Input: Global Caption: A cat sitting on a sofa near a lamp. Bounding Boxes: id: 1 | label: cat | color: #4A4A4A | region_caption: a cat | coords: [100, 200, 300, 400] id: 2 | label: sofa | color: #F5F5DC | region_caption: a sofa | coords: [0, 150, 800, 900] id: 3 | label: lamp | color: #FFD700 | region_caption: a lamp | coords: [700, 50, 850, 500] Desired Output: "A <color>#4A4A4A</color> cat is sitting on a <color>#F5F5DC</color> sofa, positioned next to a <color>#FFD700</color> lamp on the right."

For instance grounding, we used the following prompt template: Role: You are a visual instance grounding expert for precise color-control data construction. Task: Given an image and its global caption, identify caption-relevant object instances that can support reliable object-level color labeling. Return only objects that are visually present, localizable, and useful for color extraction. For each instance, return a normalized bounding box. Constraints & Rules: Caption Faithfulness: Prefer objects explicitly mentioned in the caption. You may include visually salient objects not mentioned in the caption only when they are necessary to preserve the scene context. Instance Separability: Split multiple similar objects into separate instances when they can be distinguished by position, size, or appearance. Spatial Disambiguation: Use relative positions such as "left", "right", "front", "background", "upper", or "lower" to disambiguate instances. No Hallucination: Do not output objects that are not clearly visible. Color Suitability: Prefer objects with a coherent dominant surface color. Avoid transparent, reflective, heavily shadowed, tiny, or highly textured objects when their color cannot be reliably measured. Normalized Bbox: Return each bounding box as normalized coordinates on a 0 to 1000 scale, in the order x1, y1, x2, y2. Bbox Tags: The bbox value must be wrapped with <bbox></bbox>, for example "bbox": "<bbox>[120, 85, 640, 730]</bbox>". No Coordinates in Captions: Coordinate values must appear only in the "bbox" field, not inside natural language captions or descriptions. Return Format: Return only a valid JSON list. Each item must contain "label", "instance_caption", "visual_description", "spatial_description", and "bbox". Example: [ { "label": "sky", "instance_caption": "the blue sky in the background", "visual_description": "a large coherent sky region", "spatial_description": "upper background", "bbox": "<bbox>[0, 0, 1000, 420]</bbox>" } ]

For edit-instruction refinement, we used the following prompt: You are an image edit prompt enhancement assistant. You are given: - an original edit prompt, - a source image, - an edited image, - and the target object label.

20

Your task is to enhance the original edit prompt, not rewrite it from scratch. Rules: 1. Preserve the original editing intent. 2. Always preserve the original hex color value exactly as written in the original prompt. 3. Do not replace the hex color with a natural language color name. 4. Improve wording so the prompt is natural, concise, and fully in English. 5. Keep the enhanced prompt under 30 words. 6. Mention texture, material, pattern, or shape changes only if they are clearly visible in the edited image compared with the source image. 7. Do not add extra changes unless they are visually clear. 8. If the original prompt already matches the edit well, make only minimal improvements. 9. If image quality is too poor to judge reliably, return exactly: error 10. Return only the enhanced prompt as one short sentence, or exactly: error 11. The output must contain the same hex color string from the original prompt exactly once. 12. Do not output explanations.

G

Evaluator Settings

For both ACBench-T2I and ACBench-Edit, SAM3 [15] localizes the target object region and scores are computed from the masked pixels. This differs from the original CompColor evaluator, which uses LangSAM and its released color-scoring protocol (Appendix C). For T2I, we segment the generated target object from the prompt; for editing, we segment the source object and reuse the source mask on the edited image without dilation. The predicted color is the mean RGB value inside the mask, and the final error is computed by channel-wise MAE to the target hex color. Samples with failed target localization receive zero score. This object-level evaluation setup is aligned in spirit with CompColor, while our main paper metric replaces their thresholded LAB accuracy with the continuous hex-fidelity score defined in Section 4. Object-centric evaluation with pretrained detectors or segmenters has established precedents in GenEval [30], T2I-CompBench [32], Grounded SAM [51], and TokenCompose [59]; our reliability evidence below comes from robustness sweeps and human audits rather than from the model version alone.

H

ACBench Reliability

ACBench’s primary score uses mean RGB inside the target mask followed by channel-wise MAE. Region averages can be affected by shading, material, and mask boundaries, so we test whether the paper’s ranking depends on this particular estimator, on SAM3 thresholds, or on localization failures. Region-statistic robustness. We recompute seven T2I model/configuration comparisons and three editing

models over 4,656 object instances with five estimators: Mean RGB (paper metric), MeanShift dominant RGB, Median RGB, Median Lab, and Dominant Lab (largest CIELAB k-means cluster). The alternatives cover the color-quantization and RGB/CIELAB evaluation used by ColorBind [54] and the dominant-hue Lab evaluation used by GenColorBench [12]. Table 7 reports T2I scores on the 0 to 1 scale used by the robustness script, which is the paper’s 0 to 100 convention divided by 100. Model / Configuration Ours, tgate = 0.8 Ours, tgate = 0.9 Ours, tgate = 0.0 FLUX.2-9B Base FLUX.2-4B Base Z-Image-Turbo Qwen-Image

Mean RGB

MeanShift

Median RGB

Median Lab

Dominant Lab

0.6858 0.6754 0.6383 0.3815 0.3702 0.3356 0.2521

0.6954 0.6800 0.6744 0.3734 0.3713 0.3395 0.2860

0.7341 0.7122 0.7053 0.4082 0.3886 0.3703 0.2975

0.7254 0.7176 0.6965 0.3930 0.3853 0.3546 0.2945

0.7228 0.7228 0.6931 0.4118 0.3943 0.3568 0.3003

Table 7 ACBench-T2I under alternative region-color estimators (0 to 1). The ordering is preserved for MeanShift, Median RGB, and Median Lab. Dominant Lab ties the two strongest variants at the displayed precision; every estimator preserves their advantage over the base models.

21

The same ranking holds for ACBench-Edit (Ours 0.7557 / Base 0.5890 / Qwen-Image-Edit 0.5440 under Mean RGB; all alternatives preserve the order with Spearman ρ=1.0). Relative to Mean RGB, the Ours−Base T2I gaps remain large under every estimator: +0.3156 (Mean RGB), +0.3241 (MeanShift), +0.3455 (Median RGB), +0.3401 (Median Lab), and +0.3284 (Dominant Lab). SAM3 threshold sensitivity. With all other evaluator settings fixed, we sweep the SAM3 detection threshold

from 0.1 to 0.9 and four coverage gates (min cluster, min total) ∈ {(0.15, 0.70), (0.10, 0.60), (0.05, 0.30), (0, 0)}, yielding 9 × 4 = 36 combinations. Table 8 reports the default-coverage slice. Paint-Anything ranks above FLUX.2-4B Base at every threshold, with a margin from +0.2482 to +0.3156. Across the full 36-setting grid the gap ranges from +0.2482 to +0.3308. This two-model sweep tests the stability of the performance margin and does not estimate a multi-model rank correlation. SAM3 detection threshold

Paint-Anything

FLUX.2-4B Base

Difference

0.1 to 0.5 0.6 0.7 0.8 0.9

0.6858 0.6837 0.6760 0.6623 0.5659

0.3702 0.3696 0.3642 0.3526 0.3177

+0.3156 +0.3141 +0.3118 +0.3097 +0.2482

Table 8 ACBench-T2I under SAM3 detection-threshold sweep at default coverage 15%/70% (0 to 1). Localization-failure audit. ACBench target hex values are explicit prompt specifications. Here, we audit

uncertainty from object localization. We set a SAM3 confidence threshold and score failures as zero. After verifying that high-confidence masks correctly cover the target on 20 random samples per model, we draw 600 images per model from the 1,000 ACBench-T2I outputs and measure failure rates (Table 9). Model FLUX.2-4B Base FLUX.2-9B Base Qwen-Image Z-Image-Turbo Paint-Anything

Failures / 600

Failure rate

25 14 10 1 33

4.17% 2.33% 1.67% 0.17% 5.50%

Table 9 Localization failure rates over 600 ACBench-T2I images per model.

Manual inspection attributes only 11% of failures to SAM3; most arise from the generator not producing a localizable target object. Paint-Anything has the highest failure rate among listed models, so failed localization does not give it an advantage through fewer zero-scored cases in this audit. Excluding all failures raises Paint-Anything from 68.58 to 70.93 without changing the model ranking. For ACBench-Edit, all models share the same human-verified source mask, so inter-model mask quality is matched by construction. Taken together, these checks support ACBench as a stable object-level color-fidelity diagnostic under alternative region statistics, SAM3 thresholds, and localization failures. The user study in Appendix K provides complementary aggregate preference evidence.

I

Pure-Color Anchor Resolution

Holding the remaining training recipe fixed and training for 4,000 steps, we vary only the solid-color anchor resolution (Table 10). Dropping to 256 × 256 costs 5.48 points on ACBench-T2I and 1.75 on ACBench-Edit. Raising the resolution to 1024 × 1024 adds only 0.94 and 0.46 points at a higher training cost, so we keep 512 × 512 as the default.

22

Anchor resolution

ACBench-T2I

ACBench-Edit

256 × 256 512 × 512 (default) 1024 × 1024

63.10 68.58 69.52

73.82 75.57 76.03

Table 10 Pure-color anchor resolution, with the rest of the recipe held fixed.

J

Editing under Large Color Shifts

To examine editing under different source-to-target color shifts, we analyze a subset of ACBench-Edit. We divide this subset into four equal-sized quartiles by the CIEDE2000 distance between the estimated sourceobject color and the requested target, and separately isolate near-complementary cases where both colors are chromatic and the hue rotation exceeds 120◦ . Table 11 reports scores (0 to 100) on this subset; Table 1 reports full-benchmark scores. Model Paint-Anything FLUX.2-4B

Q1 (nearest) 78.8 67.6

Q2

Q3

77.1 79.2 64.2 62.9

Q4 (farthest)

Near-comp.

88.9 56.3

86.9 50.8

Table 11 ACBench-Edit scores (0 to 100) by source-to-target CIEDE2000 distance.

Paint-Anything remains effective in every bucket and is strongest on the farthest quartile and the nearcomplementary subset, while the base model degrades as the required color shift grows. CompColor Distant under the hex-translated protocol shows the same pattern for target-pair contrast (Section 4).

K

User Study

We collect color-preference judgments from 15 participants on 80 randomly sampled ACBench-T2I prompts. Each trial shows a target swatch and two anonymized images from Paint-Anything and FLUX.2-4B. Across 15 × 80 = 1200 judgments, Paint-Anything is preferred in 55%, tied in 32%, and loses in 13%; the non-tie win rate is 55/(55 + 13) = 80.9%. Across 200 paired image-quality judgments, Paint-Anything wins 22%, ties 58%, and loses 20%, giving a 52.4% non-tie win rate. These descriptive results support the direction of the automatic color-fidelity comparison, but do not establish per-sample metric-human agreement or image-quality equivalence. Because participants and prompts recur across judgments, uncertainty estimates must account for both sources of dependence.

L

Training and Evaluation Separation

ACBench-Edit source images come from a separate internal dataset and do not overlap with the Paint-500K source images. Evaluation prompts are constructed independently and use different formats from the training captions. For the exact overlap check, benchmark prompt strings were normalized by trimming whitespace, lowercasing, and collapsing internal whitespace. The same normalization was applied to the object prompt, T2I prompt, and edit instruction fields in Paint-500K. Under this normalization, we found no exact string overlaps between benchmark prompts and training prompts.

23

M

Additional Training-Data Examples

We show 20 T2I examples and 20 editing pairs selected for visual diversity from the available training-data sample pools. The displayed text is translated from the original Chinese annotations; editing captions are shortened for readability. Target hex values are preserved. These examples illustrate the data used for supervision rather than outputs evaluated on ACBench.

A nearly black #020806 background fills most of the frame. A little left of center, a woman in a red #961E0A long dress holds several red roses in both hands and looks down at them. Red fabric spreads out below and in front of her, creating a strongly atmospheric composition.

A #E1E9EB sky stretches above distant #485152 mountains. Dark green conifers and warm yellow shrubs stand in front of the mountains. A small, pale beige calf with curly hair stands at the center, looking toward the camera. In the foreground, grass is #292E23 . A soft, warm light creates a peaceful, soothing atmosphere.

A winter forest with #BCBEBD snow in the foreground and snow-covered fallen branches scattered across the ground. #4D4A4A tree trunks fill the middle and rear of the scene, with warm sunlit highlights on some trunks. Darkness fills the deeper forest. Beams of sunlight cast shadows of varying intensity across the snow, creating a quiet, chilly woodland scene.

A naturalistic wildlife photograph in green tones. A brown rabbit sits near the lower center of the foreground, ears upright and gaze turned left. It rests on #A8B869 grass. A background of green grass and blurred vegetation surrounds the rabbit. A shallow depth of field and soft natural light create a calm, leisurely atmosphere.

A rustic still life on a wooden tabletop in warm sunlight. A pair of old sheets of paper lies in the foreground, with two yellow ears of corn near their upper center. Behind the corn, an off-white teapot sits on the left and a jar with a red fu character sticker sits on the right. A small black #050609 container stands at the lower right, beside the papers.

Figure 10 T2I training-data examples (1/4). Five image–caption examples with their target hex colors.

24

An artistic composition on a plain, neutral #898A92 background. A broken white plate in #838792 sits at the center, with fragments scattered around an apple. A red #571217 apple with yellow-orange stripes rests in the middle of the broken plate. At the center, the apple has been cut into several pieces and reassembled, with its stem still attached at the top.

Rice cake slices in #DECFC0 are stacked on a wooden plate at the center, topped with black sesame seeds, chopped nuts, and dried red fruit. At the center, the plate rests on coarse, dark brown fabric in #462A26 and #1E100E , which fills the background. An edge of a woven basket appears at the upper left. Red chilies decorate the upper and lower areas of the fabric, creating a rustic Chinese snack still life.

Outdoors, a young woman with two braids stands at the center, smiling and turning her face toward the left. Facial skin has #C6A397 and #825F4D tones. A pale pink top, visible in the lower frame, has small flutter sleeves and includes #B6B1B8 and #F6F1F2 . Blurred green plants, scattered sunlight, and fine strands of hair lifted by the wind create a fresh, soft atmosphere.

A contemporary Chinese-style portrait. On the right, a woman leans against a dark textured counter on the left, resting her left arm on it and holding a pale cyan teacup in her right hand, and she wears a black #010607 strappy long dress with a yellow-and-white printed lower garment, a fine necklace, and a tattoo on her left arm. Clusters of white flowers fill the left foreground. A beige painting with meticulous floral patterns hangs behind her in soft, elegant light.

A close-up of a white lotus in full bloom at the center, with yellow stamens, broad open petals, and several supporting green stems. A large lotus leaf in #A2CDA5 and #B9E0C4 extends over the flower at the top of the frame. More leaves and stems fill the blurred lotus-pond background, whose main tone is #ABD2A0 . Overall, the scene feels fresh and soft.

Figure 11 T2I training-data examples (2/4). Five image–caption examples with their target hex colors.

25

A nostalgic Mid-Autumn food close-up. Against a blurred background in #798275 and #535F56 , a chipped green table supports a black #080403 tray at the center. At the center, the tray holds four golden Cantonese mooncakes with baked reddish-brown patterns and stamped text such as double yolk. An edge of a dark woven food basket appears at the upper right.

A traditional Chinese portrait with low saturation and a shallow depth of field. Against a white #F5F7FA background, a woman sits sideways at the center, looking slightly downward with her hands folded in front. Black #0C0E08 hair forms a high bun with gold ornaments. An outer hanfu is dark green in #192215 and #063F43 , with subtle sleeve patterns and an orange inner layer. An elegant, quiet atmosphere fills the scene.

A painted portrait of a serious, contemplative woman at the center, looking toward the left. A mass of long golden hair in #778271 covers her shoulders. A dark top combines #403D2F and #322C20 and fills the lower part of the frame. Abstract red and gray-green brushstrokes surround her in the background, emphasizing the figure and the painting's handmade texture.

At dusk, a soft pink-purple sky in #C47B7C and #876879 fills the upper frame. Himeji Castle stands at the center, with a white main structure, several tiers of dark gray roofs with upturned eaves, and a stone foundation. Dense #141612 trees fill the foreground and surround the lower part of the castle, creating a quiet, beautiful scene with a traditional atmosphere.

A low-angle view of an ancient stone fortress in a desert landscape beneath a broad #CEDBEC sky that fills the upper half of the frame. Its weathered walls are mainly #816B5F with areas of #5C3D2C . Round towers rise above the walls, flags hang from them, and fires burn in some areas. A palm tree stands at the lower right. A sandy foreground contains stones and ruined building fragments.

Figure 12 T2I training-data examples (3/4). Five image–caption examples with their target hex colors.

26

A front-facing photograph of a fluffy cat at the center. Its fur is mainly #DAC0A6 with #BE9263 markings. It has large bright green eyes, a pink nose, and white whiskers. A small #E0E0E1 flower rests near the top of its head. At the center, the cat lies in grass among many small white flowers, against a blurred natural background mixing green and warm tones. A soft, fresh atmosphere fills the scene.

A medium close-up of an East Asian toddler with small braids at the center, looking curiously at the camera with round eyes. A softly colored painting hangs high on the #DCD6DC wall behind her, and she wears a yellow dress with small floral and cartoon bear patterns. Its broad white collar has ruffled edges and includes #C2AFA3 and #F9F1EF in the foreground.

A surrealist, ultra-wide-angle close-up in cool blue-gray tones. A woman's portrait occupies the right side, with black eyes and curly brown hair in #211E1A and #090807 covering much of that side. On the left, clouds cross a sky in #2C3B4C and #21557E . Blurred city silhouettes appear in the distant lower left, creating a mysterious atmosphere.

An oil painting with abstract expressionist and vintage influences. A woman appears in a centered, front-facing half-length portrait against a simple black #040A08 background in the upper frame, and she wears a magenta #680310 headscarf and clothing with intricate textures and colorful vertical stripes. A soft light, warm tones, and heavy yet gentle brushwork create a calm, mysterious expression and strong color contrast.

A flat forest illustration with vivid colors. At the top, the sky is #F7766B , a dark #1F0C30 tree trunk stands on the left, and the green slope at the center is #4CC795 . A person in an orange top sits at a table on the slope using a laptop. A pair of deer among the trees behind and to the right look toward the person. Dense trees in varied shades, low plants, and dark rocks surround the scene.

Figure 13 T2I training-data examples (4/4). Five image–caption examples with their target hex colors.

27

Original Image

#F9DB45

Original Image

Change the roof to #F9DB45 .

Original Image

#E7E6EF

Original Image

Change the cube to #E7E6EF .

Original Image

#81252B

Original Image

#34645F

Change the book cover to #34645F .

#CBABA4

Original Image

Change the headboard to #CBABA4 .

Original Image

#C58F94

Change the sweatshirt to #C58F94 .

Change the wallet to #81252B .

Original Image

#28361B

Change the leaves in the basket to #28361B .

#655252

Change the skirt to #655252 .

#ABC1B9

Original Image

Change the plate to #ABC1B9 .

#DB8548

Change the chest of drawers to #DB8548 .

Figure 14 Editing training-data examples (1/2). Ten source–target pairs with short instructions. Each pair places the source on the left and the target on the right.

28

Original Image

#A75747

Original Image

Change the walkway to #A75747 .

Original Image

#FA4F66

Original Image

Change the barrier belts to #FA4F66 .

Original Image

#816364

Original Image

#569FEB

Change the sky to #569FEB .

#3D3732

Original Image

Change the refrigerator to #3D3732 .

Original Image

#918C74

Change the perfume package to #918C74 .

Change the banner to #816364 .

Original Image

#012277

Change the tablecloth to #012277 .

#C07C4D

Change the umbrella to #C07C4D .

#7A7365

Original Image

Change the curtains to #7A7365 .

#BFC6D1

Change the picnic mat to #BFC6D1 .

Figure 15 Editing training-data examples (2/2). Ten source–target pairs with short instructions. Each pair places the source on the left and the target on the right.

29