FreeStyle: Free Control for Style–Content Dual-Reference Generation from Community LoRA Mining Jinghong Lan1,2∗ Wei Cheng2∗ Yunuo Chen2 Ziqi Ye1 Peng Xing2 Yixiao Fang2 Rui Wang2 Yufeng Yang2 Xuanyang Zhang2 Xianfang Zeng2 Difan Zou4 Gang Yu2‡ Chi Zhang3‡ 2
Fudan University
StepFun
Project Page
1
Community LoRA Pipeline
2
3
Westlake University
Dataset
FreeStyle Dataset & Statistics
Benchmark
4
University of Hong Kong
Weights
3
Code
FreeStyle Style Transfer and Generation Gallery
yle t St ten n Co
arXiv:2606.20506v1 [cs.CV] 18 Jun 2026
1
Filter by base models
FLUX.1
Style LoRAs
Content LoRAs
Style and Subject Driven Generation
ComfyUI Workflows
Style
Content Style+Content
Style Transfer
Figure 1. Overview of FreeStyle. 1 We collect community-created style and content LoRAs from multiple platforms and automatically compose them through standardized workflows. 2 The resulting FreeStyle dataset contains diverse style–content image triplets spanning multiple base models, artistic styles, and subject categories. 3 FreeStyle enables both style transfer and style-subject controllable image generation across a broad range of visual domains.
Abstract Style- and content-dual-reference generation aims to synthesize an image that preserves the structure and semantics of a content reference while adopting the style of a separate style reference. Despite recent progress, this setting remains challenging because models must balance content fidelity, style alignment, and instruction following while avoiding semantic leakage from the style reference. A key bottleneck is the lack of large-scale triplet data with clean content-style separation and broad long-tail style coverage. In this work, we propose FreeStyle, a scalable dual-reference generation framework based on community LoRA mining. We treat community LoRAs as compositional anchors for style and content, and design a rigorous generation and filtering pipeline to construct large-scale content–style dual-reference triplets * Equal contribution. ‡ Corresponding authors.
across multiple base models. To address content leakage, we adopt a two-stage curriculum with stage-specific disentanglement mechanisms: an attention-level enrichment constraint that suppresses style-reference leakage in the style-transfer stage, and a frequency-aware RoPE modulation strategy that targets positional-correspondence-based leakage in the harder dual-reference stage. We also introduce a benchmark covering both style-reference and dual-reference generation, with evaluations on style similarity, content preservation, aesthetics, instruction following, and VLM-based verification. The benchmark incorporates a style-invariant Content Alignment Score (CAS) and introduces a VLM-based Verification Score for evaluating generation reliability and potential cross-reference leakage. Extensive experiments show that our model achieves a strong balance among style alignment, content preservation, and leakage suppression.
1. Introduction Reference-based image generation has become an effective paradigm for controllable visual synthesis, built upon the rapid progress of diffusion models [14, 28, 29, 70] and largescale text-to-image systems [51, 57, 61, 62, 64]. Existing methods typically use external images to provide either style guidance or content reference, but style- and content-dualreference generation remains a more challenging setting: given a content reference, a separate style reference, and a text instruction, the model must synthesize an image that preserves the structure of the content reference while adopting the visual style of the style reference. This requires the model to jointly balance content fidelity, style alignment, and instruction following. A key bottleneck is the lack of large-scale triplet data with clean content-style separation and broad style coverage. Existing data construction pipelines either cover only a narrow set of styles, require costly manual curation, or produce triplets with imperfect content-style separation. In this work, we propose FreeStyle, a scalable framework for dual-reference generation based on community LoRA mining. Our key observation is that community LoRAs collectively provide a naturally curated and parameterized collection of visual concepts: each LoRA is typically trained around a coherent style, subject, or theme, and the whole community covers a broad spectrum of artistic styles and content categories. We use these LoRAs as compositional anchors for both style and content, and design a rigorous mining, generation, and filtering pipeline to construct largescale Style-Reference (SRef) and Content-Reference (CRef) triplets with broad long-tail style coverage and clean contentstyle separation. Beyond data construction, another central difficulty is content-style disentanglement. The style image often contains not only visual attributes such as color palette, texture, and brushwork, but also semantic content such as objects and layouts. When used as a conditioning signal, these semantic elements can leak into the generated image, causing unwanted hallucinations or structural distortions. We find that this leakage manifests through different mechanisms depending on the generation setting. In style-reference generation, leakage primarily arises from disproportionate attention allocation to style-reference tokens during late denoising steps. In the harder dual-reference setting, where a content reference absorbs much of the model’s attention, leakage instead occurs through local positional correspondence encoded in high-frequency RoPE components, which enables patch-level copying from the style image. To address these distinct failure modes, we adopt a twostage training curriculum with stage-specific disentanglement mechanisms. In Stage 1, the model is trained on styletransfer data to build robust style-reference generation, with an attention-level enrichment constraint that suppresses ab-
normal style-reference dominance while preserving style richness. In Stage 2, we introduce dual-reference data and apply frequency-aware RoPE modulation to the style-reference branch, suppressing high-frequency positional components that encourage copying while amplifying low-frequency components that preserve global stylistic structure. The two stages target complementary leakage pathways and together achieve a better balance among style alignment, content fidelity, and leakage control. We also introduce a benchmark for systematically evaluating both style-reference and dual-reference generation. Beyond standard feature-based metrics, we adopt the styleinvariant Content Alignment Score (CAS) from CSGO [81] to measure structural agreement after factoring out style, and propose a VLM-based Verification Score that separately quantifies style-transfer reliability and content preservation under potential cross-reference leakage. This evaluation protocol exposes the trade-offs that a single aggregate score would obscure. Our contributions are summarized as follows: • We propose FreeStyle, a scalable dual-reference generation framework that mines community LoRAs as compositional anchors for style and content, enabling large-scale construction of SRef and CRef triplets across multiple base models with broad long-tail style coverage. • We introduce a systematic benchmark for both stylereference and dual-reference generation, together with a VLM-based Verification Score that separately evaluates style-transfer reliability and content preservation under potential cross-reference leakage. • We propose a two-stage training strategy with stagespecific disentanglement: an attention-level enrichment constraint for style-reference generation, and frequencyaware RoPE modulation for dual-reference generation, each targeting a distinct leakage mechanism.
2. Related Work 2.1. Reference-Based Generation and Stylization Reference-based generation controls image synthesis using external visual examples. For content control, ControlNet [84] and T2I-Adapter [50] inject spatial signals such as edges, depth, or poses into diffusion models; IP-Adapter [82] uses decoupled cross-attention for image-reference conditioning; and personalization methods adapt models to userprovided concepts via fine-tuning, low-rank adaptation, or token optimization [21, 30, 37, 63, 68]. Style-reference generation further requires extracting visual attributes—color, texture, brushwork—from a reference image. Neural style transfer has evolved from optimizationbased and feed-forward formulations [16, 23, 33, 73] through arbitrary zero-shot stylization [9, 31, 42, 67] to attention- and transformer-based methods [13, 45, 54], re-
versible flows [2], learned linear transforms [41], featuredistribution and optimal-transport matching [36, 86], contrastive objectives [8, 80], and wavelet-based designs [47, 83]. More recently, diffusion models enable trainingfree stylization via shared or swapped attention [27, 32], inversion-based methods [10, 87], and LoRA merging [19, 66], while GAN-based generators [24, 34, 35, 88] offer domain-specific style control. However, a style reference often carries semantic content that leaks into the output. StyleAlign [27], InstantStyle [76], DEADiff [58], and CleanStyle [18] address this through disentanglement or purification strategies; CSGO [81] proposes end-to-end contentstyle composition; and EasyRef [89] supports generalized group-image references via multimodal LLMs. These methods advance individual axes of reference-based generation, but most do not target large-scale supervision for explicit content-style dual-reference generation.
2.2. Content and Style Dual-Reference Generation In the dual-reference setting, the model must simultaneously preserve the content reference, adopt the style reference, and follow a text instruction—requiring multi-image fusion and content-style disentanglement beyond single-reference or text-guided editing methods [6, 26, 48]. General-purpose editors such as GPT-Image 1.5 [52], Nano Banana Pro [25], Qwen-Image-Edit [59], FLUX.2 [klein] [5] and iMontage [20] handle multi-image inputs but are not optimized for this setting and can be unstable when content preservation, style alignment, and instruction following must all be satisfied simultaneously. Among task-specific methods, USO [79] constructs compositional content-style triplets through subjectdriven generation and de-stylization, representing an important step toward dual-reference control. Nevertheless, the field still lacks large-scale, diverse, and cleanly separated triplet data, as well as systematic benchmarks evaluating the trade-offs among content fidelity, style alignment, instruction following, and leakage suppression.
2.3. Data Construction for Controllable Generation High-quality triplet data consisting of content reference, style reference, and target image with clean separation, is fundamental to dual-reference generation, yet existing pipelines remain limited in scale, diversity, or separation quality. Preference-based datasets such as Premier [78] mine user interactions but are not designed for contentstyle disentanglement. Synthetic stylization methods such as MegaStyle [22] and OmniStyle [77] scale more easily but inherit the style range and artifacts of the underlying generator. Hybrid pipelines such as TeleStyle [85] improve quality through manual curation yet are difficult to scale to broad style categories. De-stylization-based methods such as USO [79] directly target compositional supervision but may weaken structural details during content recovery. In
contrast, we mine community LoRAs as scalable compositional anchors for style and content, constructing large-scale triplets with broad style diversity and clean content-style separation across multiple base models.
3. Method Overview We present FreeStyle, a framework for style- and contentdual-reference image generation built on three tightly coupled components. Data (§4). We construct two complementary datasets. A style-transfer pipeline (§4.1) generates triplets by applying controlled stylization to diverse content images, providing clean supervision for basic style-reference generation. A community-LoRA mining pipeline (§4.2), the central data contribution of this work, treats LoRA weights as compositional anchors for both style and content, enabling large-scale dual-reference triplet construction with broad long-tail style coverage. Training (§5). We adopt a two-stage curriculum progressing from style-reference generation (Stage 1, trained on styletransfer data) to the harder dual-reference setting (Stage 2, mixing LoRA-mined triplets with style-transfer data). Each stage faces a distinct content-leakage mechanism and employs a corresponding disentanglement strategy: an attentionlevel enrichment constraint (§5.1) for Stage 1 and frequencyaware RoPE modulation (§5.2) for Stage 2. Benchmark (§6). We introduce an open benchmark for both style-reference and dual-reference generation. It adopts the style-invariant Content Alignment Score (CAS) from CSGO [81] for content evaluation and introduces a VLM-based Verification Score for measuring style-transfer and content-preservation reliability under potential crossreference leakage.
4. Data Pipeline To support two-stage training, we construct two complementary datasets. We first describe the style-transfer data that provides supervision for Stage 1 (§4.1), and then present the community-LoRA-based mining pipeline that produces large-scale dual-reference triplets for Stage 2 (§4.2).
4.1. Style-Transfer Data Construction Stage 1 requires large-scale style-transfer triplets (content image, style reference, stylized target) with clean separation between content structure and visual style. We construct this dataset by leveraging the state-of-the-art generative model Nano Banana Pro [25] combined with bilateral consistency filtering, as illustrated in Figure 2.
1. Source Image Collection Collect from website
2. Style Transfer Validated Trigger Word List American Tattoo Art transfer into <trigger> style
Diverse keywords: Animal, human, scene...
Gemini-3
4. Triplet Construction
3. Feature filtering
Valid Style Transfer Triplets
Content Alignment 0.87 0.92
DINO-v2
0.13
Style Similarity Style Encoder
Content
Style
Target
Figure 2. Overview of the Style-transfer Data Construction Pipeline. (1) Content collection. We crawl a large set of raw content images from diverse websites, covering categories such as landscapes, human subjects, and everyday objects. (2) Triggerword stylization. We validate a community style-trigger-word list and retain roughly 600 stable triggers, then stylize each content image through a fixed “transfer into [trigger] style” template. (3) Bilateral consistency filtering. For every stylized output we verify content fidelity against its source image via DINOv2 feature similarity, and measure style similarity with the ONEIG image encoder. Samples that pass both checks are assembled into clean style-transfer triplets.
Specifically, we first curate diverse content images from the web spanning landscapes, human subjects, and everyday objects, and validate approximately 645 style trigger words for generation stability. Each content image is then stylized through a fixed prompt template (e.g., “transfer into [trigger] style”). We apply bilateral consistency filtering to every output: content fidelity is verified against the source via DINOv2 [53] feature similarity, and style consistency is measured against anchor style images using the ONEIG style encoder. Only samples passing both checks are retained. When assembling triplets, the style reference is sampled from a different content source to ensure contentstyle independence, and text prompts are drawn from a GPT [1]-generated pool of style-transfer instructions.
4.2. Community LoRA Mining and Triplet Construction Thanks to the continuous evolution of ComfyUI [12], we observe that the open-source community hosts a massive amount of high-quality LoRA models and creative workflows. These LoRAs are exceptionally rich and diverse, covering an extensive range of content categories including characters, architecture, scenes, animals, daily objects, food, and vehicles, as well as nearly all possible stylistic categories. However, constrained by factors such as the intrinsic instability of certain LoRA weights, the potential of these open-source resources has not been fully exploited. To
this end, we mine community LoRAs from platforms like Civitai [11], TensorArt [72], and Liblib [43], and design a highly robust pipeline for filtering and image generation, ultimately constructing a large-scale dataset for dual-reference generation. We first crawl the LoRA weights along with their web metadata. Categorizing them by their base models, we select LoRA weights built upon three well-established textto-image backbones: Illustrious [55], FLUX-dev [39], and Qwen-Image [59]. For each base model, we carefully design and tune specific ComfyUI workflows. To fully unleash the effectiveness of these LoRAs while ensuring diversity, we draw on community-summarized heuristics to design vocabularies that are highly compatible with both style and content generation. Furthermore, drawing inspiration from the category taxonomy of the OpenImages dataset [38], we incorporate extensive object-related vocabularies to enhance the content richness of the generated style reference images. Based on these vocabularies, we sample and construct a massive prompt pool that allows for diverse random prompt generation. With these preparations, we execute the following steps to obtain rich triplet images, as illustrated in the overall pipeline in Figure 3. 1. Collection and Filtering of Stable LoRA Weights. To ensure a high success rate for subsequent combinations, we must first screen out reliable style and content LoRAs. To make the generation compatible with both types of LoRAs, the prompt pool in this stage primarily consists of scenerelated vocabulary combined with the inherent trigger words of each LoRA. We generate a 3 × 3 grid of 9 preview images for human experts to evaluate whether the generation is stable and to categorize the LoRA as either content-oriented or style-oriented. A LoRA is considered stably triggered only if at least 7 out of the 9 images in the grid exhibit consistent quality. Once the stable content and style LoRAs are obtained, given their sheer volume, we perform aesthetic scoring [75] and ranking on the content LoRAs to preserve content diversity and reduce the number of candidate combinations. We then apply non-uniform sampling based on these scores: higher-scoring content LoRAs have a higher probability of being sampled, thereby reducing the effective scale of the LoRA pool. Figure 4 illustrates the Sankey diagram of our LoRA filtering process and the final distribution of curated LoRAs, where the style categorization aligns with the taxonomy used in our benchmark. 2. Generation of Rich and Effective Reference Images. With the stable LoRAs secured, we use them to generate reference images that are both content-rich and highly representative of the LoRAs’ own themes. Here, we utilize the comprehensive prompt pool prepared initially, which contains approximately 40k distinct prompt combinations, and
Single-Lora Models Filtering Community Sources
Single-Lora Results Filtering Valid LoRA Model IDs
Website Crawling
Demo Images, Trigger Word, Base Model, Checkpoint, etc.
Lora Lora Lora
Lora Lora Lora
Dual-Lora Pairs Filtering Dual-LoRA Combo Filtering
……
LoRA Detail Page Analysis
Text Prompt Pool
Trigger Words
ComfyUI Workflow Pipeline
Meta Info
Trigger Words
Lora
Pre-filtered by VLMs
Prompt
Base
VAE
Noise
Denoise
Save Img
Qwen3-VL
Get lora theme, choose proper words type
Human Expert Filtering
Valid LoRA List
ComfyUI Workflow Pipeline Content LoRAs, Style LoRAs
……
Content LoRA Images
LoraC
Text Prompt Pool
Trigger Words bad case rate criteria
LoraB
Load Img
Demo Images with Prompts
LoraC
LoraB
LoraA
Content LoRAs, Style LoRAs
LoraA
Valid Content/Style Lora Models
LoraB
LoraB
LoraC
…… ComfyUI Workflow Pipeline
Aesthetics Expert Filtering
Lora Lora Lora
LoraB
LoraC
Valid LoRA Combos
Style LoRA Images
Lora Lora Lora
LoraA
Content LoRAs
Is Content/Style Consistency?
Y/N Valid Ref Images
Valid Reference Image
Is Content & Style Consistency?
Style LoRAs
Y/N
Pairing
Valid Images
Valid Triplet
Valid Content Style Triplets
Figure 3. Overview of the FreeStyle Data Construction Pipeline. 1 Single-LoRA Models Filtering. Community content and style LoRAs are collected and curated to build a high-quality LoRA repository. 2 Single-LoRA Results Filtering. Representative reference images are identified through metadata analysis and generation-based validation. 3 Dual-LoRA Pairs Filtering. Compatible content–style LoRA pairs are selected and combined to construct high-quality content–style triplets.
combine them with the LoRAs’ trigger words. At least 20 images are generated for each LoRA. Since the generated outputs are not always perfectly stable, we use the preview images collected from the web as references and employ Qwen3-VL [3] for verification. The verification adopts the binary VLM judgment and majority-voting protocol used in our benchmark, yielding accurate and stable filtering results. Notably, we design distinct parameters and workflows for different base models to perform batch generation via ComfyUI services; these workflows will also be open-sourced. 3. Filtering of Valid LoRA Combinations. LoRAs can be combined with one another. In many excellent community workflows, multiple LoRAs are often applied simultaneously to a single base model using various advanced techniques (e.g., community experience suggests that inserting LoRAs into different layers of SDXL [57] yields diverse effects). Although style and content LoRAs do not inherently conflict in their domains, the combined effect largely depends on the stability of the LoRAs themselves. However, we noticed that the content of some LoRAs carries intrinsic stylistic traits, leading to degradation and instability during actual combination. In practice, this instability is persistent: if a combination has inherent conflicts, batch testing with even
highly diverse prompts will still yield poor results. Therefore, LoRA combinations must be rigorously filtered. To address this, we perform a Cartesian product of the style and content LoRAs, generate one image per combination, and conduct bilateral content and style verification against the single-LoRA images generated previously. In this step, the individual filtering success rate is approximately 0.6 for style and 0.8 for content, while the success rate for the bilateral combination drops to about 0.4. Through this rigorous process, we acquire a massive set of stable and reliable LoRA combinations. 4. Batch Generation of Target Triplet Data. Having established stable LoRA combinations, we leverage the rich vocabulary data to batch-generate massive target images for our triplets. These target images are again subjected to bilateral consistency verification against the corresponding single-LoRA style and content reference images. Importantly, because the validity of the LoRA combinations has already been screened, the verification success rate during batch generation reaches as high as 0.8, dramatically boosting efficiency and stability. Ultimately, we successfully generated and filtered 273k triplets using the FLUX model, 33k using the Qwen model, and 172k using the Illustrious
counts shown on labels
a. Community LoRA mining flow POOL
b. Curated LoRA categories
link width ~ sqrt(count)
BASE
TYPE
OUTCOME
SAMPLE
FLUX unstable 15.4k
FLUX base 18.0k
Other model
FLUX content
13.0k
912
sampled 100
FLUX style 1.7k
Collected LoRA pool 68.6k
7.0k
800
Illus. style 210
Illus. unstable 10.8k 636
Characters People Vehicles Animals Clothing Scenes Others
910 288 52 38 31 27 12 532
Advertising illustration
sampled
Qwen split content 29 | style 58 | unstable 549
Style
Style categories
Illustrious base 18.8k Illus. content
Qwen base
Content
Content categories
Qwen unstable 549
2D Japanese anime Line art Game art Modern art design 3D anime art Traditional western art 2.5D style 2D Western animation Traditional eastern art Material art
333 250 141 139 137 132 91 83 49 22 10
30
100
300
1000
frequency count (log scale)
Figure 4. Statistics of Community LoRA Mining and Filtering. The Sankey diagram (left) traces the successive stages of our LoRAfiltering pipeline: raw community LoRAs first undergo stability screening, and, to curb the combinatorial explosion of style–content pairings, we additionally sub-sample the content LoRAs via aesthetic-score-weighted non-uniform sampling, where higher-rated LoRAs are retained with higher probability, yielding the final set of curated weights. The distribution (right) classifies these final LoRAs by theme, following the same taxonomy as our soon-to-be-released benchmark.
model. The resulting dataset is exceptionally massive and highly diverse in its compositional combinations. Finally, we employ a Qwen3-VL model [3] to generate the image content prompts that guide the model in bridging the content reference image and the target image. Through these meticulous steps, we assemble a comprehensive and highly diverse triplet dataset consisting of the content reference image, the style reference image, the generative prompt, and the target image.
5. Style-Content Disentanglement Training Our method introduces no additional architectural modifications to the base model. We inject text, contentreference, and style-reference conditions into the MMDiT blocks [15, 17, 56, 74], following the multi-image input paradigm of Qwen-Image-Edit, whose text branch is encoded by a vision-language model [4] and whose generator is trained with a flow-matching objective [44, 46]. We adopt a two-stage training curriculum. In Stage 1, the model is trained exclusively on style-transfer data (Sec. 4.1) to build robust style-reference generation capability. At this stage, content leakage—where semantic elements from the style reference are inadvertently copied into the generated image—primarily manifests as disproportionate attention allocation to style-reference tokens during late denoising
steps. We address this with an attention-level enrichment constraint (Sec. 5.1). In Stage 2, we mix community-LoRAmined dual-reference triplets (Sec. 4.2) with style-transfer data, introducing the harder setting where the model must simultaneously respect a content reference, a style reference, and a text instruction. Now, the content reference absorbs a substantial share of the model’s attention, and the stylereference attention no longer exhibits the same pronounced asymmetry seen in Stage 1. Yet content leakage persists through a different channel: local positional correspondence encoded by high-frequency RoPE components, which enables patch-level copying from the style image. We address this with frequency-aware RoPE modulation (Sec. 5.2). The two stages thus target distinct leakage mechanisms and employ complementary disentanglement strategies.
5.1. Attention Constraint for Style-Reference Generation Observation: content leakage correlates with attention asymmetry. Before designing any constraint, we first analyze how attention allocation over different semantic groups correlates with content leakage in style-reference generation. Figure 5 presents representative evidence. In the attention maps (left panel), the vertical axis corresponds to queryside noisy-latent tokens and the horizontal axis to semantic partitions on the key side, with the dark-red and dark-blue re-
gions indicating content-reference and style-reference token ranges, respectively. Comparing leakage cases against successful transfers reveals a clear pattern: leakage cases exhibit substantially broader and more persistent high-response bands over the style-reference region, whereas successful cases maintain more compact and stable responses. The right panel further shows the style-reference attention mass ratio across denoising time in first transformer block, confirming that the asymmetry intensifies in late denoising steps and is most prominent at the first transformer block, where the global semantic layout of the generated image is determined. This observation reveals that content leakage is not a static model property, but a time-varying failure pattern that can be localized along both the denoising and depth axes. It motivates us to (i) define a quantitative metric that captures this disproportionate allocation, and (ii) convert it into a differentiable constraint. Group-wise attention enrichment. To quantify the above phenomenon, we measure how much attention each semantic group of tokens receives relative to its size. Let A(t,ℓ) be the attention map at denoising step t and transformer block ℓ, with entry Aqk the attention weight from a noisy-latent query q to a key token k. We split the keys into semantic groups—text, content reference (cref), and style reference (sref)—writing Gg for group g and K for all keys. The share of attention that group g receives is P P (t,ℓ) q k∈G Aqk (t,ℓ) pg = P P g (t,ℓ) , (1) q k∈K Aqk A larger group naturally attracts more attention, so we normalize by the group’s size fraction to obtain the enrichment score: (t,ℓ) pg . (2) Eg(t,ℓ) = |Gg |/|K| Here Eg = 1 means group g is attended to exactly in proportion to its size; Eg > 1 indicates enrichment (the group draws more attention than its size warrants); and Eg < 1 indicates suppression. Because the score factors out group size, it places groups of different sizes on equal footing while still resolving how attention shifts along both the denoising and depth axes. Attention-regularization losses. Based on the above analysis, we convert the enrichment metric into two lightweight regularizers. Both are applied only at the first transformer block (ℓ = 0), where attention over the style reference is most predictive of leakage. The two losses share a single two-sided squared hinge that keeps a quantity x within a target band [a, b]: ψ[a,b] (x) = max(0, a − x)2 + max(0, x − b)2 .
(3)
Enrichment loss. For each noisy-latent query q, the style-reference enrichment at the first block follows Eq. (2) restricted to that query, Esref (t, q) = P (t,1) We keep it inside |Gsref |/|K| . k∈Gsref Aqk [αlo , αhi ] and weight later denoising steps more heavily by (1 − t): h i Lenrich = Et, q (1 − t) ψ[αlo , αhi ] Esref (t, q) . (4) We set αlo = 0 and αhi = 0.6: the zero lower bound means we never force attention onto the style reference and only penalize over-attention beyond 0.6× the size-matched baseline, which is what drives content copying. The (1 − t) weight concentrates the constraint on late denoising steps, where reference influence actually materializes. Entropy loss. Suppressing over-attention can collapse the style-reference attention onto a few keys and weaken style transfer. We therefore also regularize the shape of each (t,1) P (t,1) query’s distribution Ãqk = Aqk / k′ ∈Gsref Aqk′ over Gsref through its normalized entropy Ĥq = −
X 1 Ãqk log Ãqk ∈ [0, 1], log |Gsref |
(5)
k∈Gsref
keeping it inside a narrow band with the same hinge: h i Lent = Et, q ψ[βlo , βhi ] Ĥq .
(6)
We use βlo = 0.06 and βhi = 0.14: below the band the attention collapses onto a few style tokens, above it the attention smears out uniformly and loses focus. The full objective augments the flow-matching diffusion loss Ldiff with both regularizers: L = Ldiff + λe Lenrich + λh Lent ,
(7)
where λe = λh = 0.1 so the two terms gently shape attention without overpowering the main objective. Together they keep content leakage under control while preserving style diversity.
5.2. Frequency-Aware RoPE Modulation for DualReference Generation After Stage 1, the model has acquired robust style-reference generation with content leakage well controlled by the attention constraint. However, when we transition to the dualreference setting in Stage 2, the nature of the task fundamentally changes: the model must now attend to a content reference, a style reference, and a text prompt simultaneously, using the content reference as the structural scaffold while transferring style from the style reference. In this jointconditioning regime, the content reference absorbs a large share of the model’s attention budget, and the style-reference
Content Leakage
Step 1
Mass Ratio of Style Token
FreeStyle Model
Style-Content Entanglement Analysis Step 10
Transfer the style of Image 1 to Image 2 VAE Encoder + Patchify
Qwen2.5-VL
Qwen-Image-Edit MMDiT Block
No Leakage
Style Tokens in 1st Block by Steps
Attention
×N
Style
Content
Style+Content
Content Token
Flow Matching Loss
Style Token
Style Enrichment Loss
Figure 5. Analyzing Content Leakage through the Lens of Attention. Left: failure cases exhibiting semantic content leakage from the style reference, contrasted with successful leakage-free transfers, where leakage cases produce broader and more persistent high-response bands than successful transfers. This observation motivates our group-wise attention-enrichment constraint. Middle: our dual-reference architecture together with the style-reference attention mass-ratio dynamics across denoising time in the first transformer block. Right: Attention mass ratio evaluation across different time steps in first block, the orange curve denotes the variant without the style-enrichment loss, whereas the green curve denotes the full model.
enrichment score no longer exhibits the pronounced latestage spike observed in the style-only setting. As a result, the attention-based enrichment constraint from Sec. 5.1 loses much of its discriminative power. Yet content leakage from the style reference persists—but through a different mechanism. We hypothesize that the leakage pathway shifts from attention magnitude to positional correspondence: the high-frequency components of RoPE [71] encode fine-grained spatial relationships between tokens, which can enable the model to establish patch-level copying mappings between the style reference and the generated output. Inspired by the frequency-aware modulation strategy in [49], we regularize the rotary positional embeddings of the style-reference branch. The core idea is to suppress high-frequency RoPE components that encourage local copying, while amplifying low-frequency components that better preserve global stylistic structure. Concretely, let d ∈ {0, . . . , D/2 − 1} index the twodimensional RoPE chunks and let D denote the RoPE dimensionality. We assign a smooth frequency-dependent scale to the sref keys: sd = shf + slf − shf
d D/2 − 1
β ,
(8)
where shf < 1 suppresses high-frequency locality, slf > 1 amplifies low-frequency global guidance, and β controls the smoothness of the interpolation across frequency bands. These scaling factors are fixed throughout the denoising process and do not depend on the timestep. We apply this modulation only to the style-reference branch, while leaving the content-reference branch unchanged.
6. Benchmark Existing evaluations of style-reference and dual-reference generation are fragmented: different works report different metrics on private test sets, making cross-method comparison unreliable. To address this, we introduce an open benchmark with a fixed reference set, standardized prompts, and a mixed evaluation protocol that separates style fidelity, content preservation, and leakage control into distinct axes rather than collapsing them into a single score. We will publicly release the benchmark, prompts, and evaluation code.
6.1. Benchmark Construction The benchmark is built from 200 content reference images and 200 style reference images, curated to maximize diversity. Content images span a broad range of object categories, scenes, and compositions, while style images cover artistic domains from oil painting and watercolor to 3D rendering, pixel art, and abstract art. Reference pairings are sampled without replacement to prevent any single style or content from dominating the evaluation. For each pairing, the text prompt is drawn from a GPTgenerated [1] prompt pool of style-transfer-oriented instructions. We generate multiple prompt variants per style category to reduce prompt-specific bias. The benchmark covers two evaluation settings: • Style-reference generation (SRef): given a source image, a style-reference image, and a text instruction, generate an image that preserves the source content while adopting the visual style of the style reference. • Content-style dual-reference generation (CRef+SRef): given a content reference image, a style reference image, and a text prompt, generate an image that preserves the
structure of the content reference, transfers the style of the style reference, and respects the text instruction.
6.2. Evaluation Metrics We adopt a mixed evaluation protocol combining featurebased encoder similarities, VLM-based judgments, and aesthetic scores. Each evaluation axis (style fidelity, content fidelity, instruction following) includes both an encoder-based metric and a VLM-based metric, since the two capture complementary aspects: encoder similarities measure continuous perceptual distance, while VLM judgments assess higherlevel semantic alignment that single-axis feature distances may not fully reflect. All VLM-based scores (VLM-S, VLMC, VLM-F, Ver-S, Ver-C) are computed using Qwen3-VL [3] as the judge. Style-fidelity axis. ONEIG [7] provides a perceptual estimate of stylistic consistency. CSD [69] measures style similarity using a contrastive descriptor trained to separate style from content. VLM-S (VLM-Style) asks a visionlanguage model to rate how well the generated image adopts the overall visual style of the reference, yielding a scalar score. Content-fidelity axis. DINOv2 [53] reports cosine similarity between the generated image and the content reference in the self-supervised DINOv2 embedding space, sensitive to object layout and semantic structure. CAS (Content Alignment Score), adopted from CSGO [81] and detailed below, measures structural agreement after removing style-carrying channel statistics, thereby reducing the influence of appearance differences on content evaluation. VLM-C (VLMContent) asks a vision-language model to rate how well the generated image preserves the structure and semantics of the content reference. Instruction-following axis (CRef+SRef only). CLIPT [60] reports image-text cosine similarity in CLIP space. VLM-F (VLM-Follow) asks a vision-language model to judge whether the generated image obeys the text instruction, yielding a scalar score. VLM verification. Ver-S and Ver-C (defined below) measure the proportions of generated images that are verified as successful in style transfer and content preservation, respectively, using repeated binary VLM judgments with majority voting. Aesthetics. LAION-Aesthetic [40] and V2.5-Aesthetic score [75] the raw visual appeal of each output, guarding against degenerate solutions that satisfy similarity metrics but look unnatural.
Content Alignment Score (CAS). We adopt the Content Alignment Score proposed in CSGO [81] to evaluate content preservation while reducing sensitivity to stylistic appearance. CAS builds on the observation that channel-wise feature statistics capture appearance-related information, whereas instance-normalized features retain more of the underlying spatial structure [31]. Specifically, let ϕ(·) ∈ RL×C denote the DINOv2 [53] patch-token features of an image resized to 512 × 512, where L is the number of tokens and C the channel dimension. For each channel c, we compute its mean and standard deviation over the L tokens, L
1X µc (I) = ϕ(I)l,c , L l=1 v u L u1 X 2 σc (I) = t ϕ(I)l,c − µc (I) + ϵ, L
(9)
l=1
and apply instance normalization to remove the stylecarrying statistics: ϕ̂(I)l,c =
ϕ(I)l,c − µc (I) . σc (I)
(10)
Given a generated image Ig and its content reference Ic , CAS is the mean squared error between their style-normalized features: L
CAS =
C
2 1 XX ϕ̂(Ig )l,c − ϕ̂(Ic )l,c . LC c=1
(11)
l=1
A lower CAS indicates better content preservation, since the metric isolates structural agreement after factoring out style differences. Verification Score. Feature-based metrics measure average similarity but do not directly indicate whether individual generations reliably satisfy the desired style-transfer and content-preservation criteria. To complement them, we introduce a VLM-based Verification Score that aggregates repeated binary judgments. Given the token log-probabilities ℓ0 and ℓ1 assigned by the VLM to the outputs “0” and “1”, respectively, we normalize them over the two candidate labels: exp(ℓc ) pc = , c ∈ {0, 1}. (12) exp(ℓ0 ) + exp(ℓ1 ) The predicted binary label is ŷ = arg max pc .
(13)
c∈{0,1}
We query the VLM three times for each generated–reference image pair. A pair is counted as a successful match if at least two of the three judgments predict ŷ = 1. Because each generated image is evaluated along two independent axes, we report two Verification Scores:
Table 1. Quantitative Comparison on the Style-reference (SRef) Benchmark. Metrics are grouped by evaluation axis: style fidelity (feature-based and VLM-based), content fidelity, VLM-based verification, and aesthetics. Closed-source models are shown above the rule for reference only. Among open-source methods, the three purple shades mark the top-3 results per metric, with our model at the bottom. ↑: higher is better; ↓: lower is better. Style
Method ONEIG↑
CSD↑
Content
VLM Verification
Aesthetics
VLM-S↑
DINO↑
CAS↓
VLM-C↑
Ver-S↑
Ver-C↑
LAION↑
V2.5↑
Closed-source commercial models Gemini 0.416 0.619 Seedream 0.441 0.636
7.848 7.023
0.840 0.836
1.452 1.077
8.798 9.154
0.554 0.334
0.852 0.905
6.157 6.495
5.749 5.923
Open-source methods USO 0.542 CSGO 0.520 EasyRef 0.270 0.490 FLUX.2 [klein] TeleStyle 0.459 Qwen-Image-Edit 0.290 OmniStyle 0.389 Ours 0.468
3.744 6.193 2.073 6.068 5.555 3.487 6.247 7.142
0.808 0.652 0.612 0.811 0.859 0.865 0.757 0.809
1.282 1.656 1.904 1.084 0.924 0.918 1.759 1.175
9.237 1.516 0.428 8.335 9.566 9.155 6.247 8.919
0.382 0.038 0.000 0.285 0.449 0.261 0.329 0.482
0.928 0.715 0.129 0.532 0.961 0.910 0.483 0.928
5.971 5.430 5.403 6.682 6.173 6.131 4.901 6.302
5.575 4.676 4.415 6.018 5.470 5.657 4.542 5.709
0.531 0.665 0.578 0.655 0.613 0.589 0.603 0.639
Table 2. Quantitative Comparison on the Dual-reference (CRef+SRef) Benchmark. Metrics are grouped by evaluation axis. Instructionfollowing metrics (CLIP-T and VLM-F) are grouped together since both measure prompt adherence. Notation and shading conventions follow Table 1. Style
Method ONEIG↑
CSD↑
Content
Instruction
VLM Verification
Aesthetics
VLM-S↑
DINO↑
CAS↓
VLM-C↑
CLIP-T↑
VLM-F↑
Ver-S↑
Ver-C↑
LAION↑
V2.5↑
Closed-source commercial models Gemini 0.369 0.590 Seedream 0.372 0.601
5.113 5.621
0.749 0.759
1.723 1.607
7.447 7.723
0.314 0.319
9.492 9.565
0.479 0.522
0.411 0.477
6.740 6.969
6.135 5.593
Open-source methods 0.450 USO Qwen-Image-Edit 0.232 FLUX.2 [klein] 0.268 TeleStyle 0.362 Ours 0.387
3.753 2.386 2.388 4.251 5.467
0.809 0.723 0.758 0.760 0.739
1.283 1.731 1.702 1.590 1.639
9.139 6.141 7.034 7.338 7.038
0.245 0.320 0.327 0.302 0.308
2.825 9.220 9.432 7.994 8.909
0.390 0.118 0.124 0.315 0.409
0.916 0.320 0.361 0.550 0.462
5.943 6.587 6.833 6.800 6.747
5.566 5.699 5.734 5.774 5.643
0.246 0.489 0.602 0.585 0.575
• Style Verification Score (Ver-S): measures the proportion of generated images judged to correctly transfer the target style. • Content Verification Score (Ver-C): measures the proportion of generated images judged to preserve the content reference under potential semantic leakage from the style image. Higher values indicate more reliable style transfer and content preservation. The Verification Score complements feature-based similarities by providing a binary pass/fail signal, offering an additional perspective on whether the model successfully balances style transfer and content preservation.
7. Experiments We evaluate FreeStyle on the benchmark introduced in Sec. 6 under both the style-reference (SRef) and dual-reference (CRef+SRef) settings. All metrics are defined in Sec. 6.2; ↑ denotes higher-is-better and ↓ denotes lower-is-better. For closed-source commercial baselines, we report the latest pub-
licly available versions at the time of evaluation; in particular, the Seedream [65] entries correspond to Seedream 4.5, queried through its official API under the same reference images and prompts as all other methods.
7.1. Implementation Details We use the Qwen-Image-Edit-2511 [59] backbone and train with a flow-matching objective. Stage 1 is trained on styletransfer data for 120k steps with batch size 32 and learning rate 6.0e-06. Stage 2 mixes community-LoRA-mined triplets with style-transfer data (ratio 3:1) and trains for an additional 24k steps. Both stages use 8× H100 GPUs. The attention-enrichment constraint is applied from the beginning of Stage 1 with λe = λh = 0.1. Frequency-aware RoPE modulation is introduced at the start of Stage 2 with β = 2, shf is 0.9, and slf is 1.2, following the experimental setup in [49].
7.2. Benchmark Results Tables 1 and 2 report the SRef and CRef+SRef results, respectively. We highlight several observations. Feature-based and VLM-based metrics should be evaluated jointly. A natural trade-off exists between contentoriented and style-oriented metrics in this task. Models with conservative stylization may obtain high DINOv2 and CAS scores while achieving limited style transfer, whereas models with stronger stylization may obtain high ONEIG or CSD scores at the cost of content preservation or increased reference leakage. For example, CSGO achieves the highest CSD score (0.665), while its relatively low VLM-Content (1.516) and Ver-C (0.715) suggest that the high feature-based style similarity does not necessarily correspond to reliable content preservation. Conversely, Qwen-Image-Edit performs best on DINOv2 (0.865) and CAS (0.918), but its lower VLM-Style score (3.487) suggests comparatively limited style transfer. These results illustrate that feature-based metrics alone do not fully distinguish successful style transfer from imbalanced content-style trade-offs; the feature-based metrics on each axis should therefore be read together with the corresponding VLM judgment and Verification Score. FreeStyle achieves the best overall balance. On the SRef benchmark, our method ranks first on VLM-Style (7.142) and Ver-S (0.482), indicating the strongest VLM-verified style transfer among open-source methods, while maintaining competitive content preservation (Ver-C second at 0.928) and aesthetics (LAION second, V2.5 second). On the harder CRef+SRef benchmark, the same pattern holds: FreeStyle ranks first on VLM-Style (5.467) and Ver-S (0.409), with competitive performance across all other axes. No single baseline achieves comparable balance: USO excels on content metrics but scores poorly on style and instruction following; FLUX.2 [klein] leads on instruction metrics but transfers style weakly (VLM-Style 2.388). Qualitative comparison. Figure 6 shows that our method captures the target style (brushwork, texture, palette) with greater precision while preserving the content structure. Other baselines either introduce structural artifacts or inadvertently copy semantic content from the style reference. Figure 7 extends this comparison to the dual-reference setting, where our method exhibits substantially less semantic leakage than all competing baselines under the same contentstyle-prompt conditioning.
7.3. Ablation Studies We validate the three core design choices of FreeStyle: the attention-map constraint, the frequency-aware RoPE modulation, and the training data pipeline. All ablation variants
share the same backbone and training budget, differing only in the component being tested. Table 3 reports VLM-based leakage scores (0–10 scale, lower is better) for the first two components. Table 3. VLM Leakage Score Ablation (0–10, lower is better). Each row isolates one component by comparing the full model against a variant without it. Component
w/o
w/
Enrichment loss (SRef) RoPE modulation (CRef+SRef)
2.674 1.047
0.522 0.453
Attention-map constraint. Figure 8 illustrates the effect of removing the enrichment loss during Stage 1 training. Without the constraint, semantic entities from the style reference are frequently hallucinated in the generated image (e.g., objects or architectural elements from the style image appear in the target scene). Enabling the constraint eliminates this leakage while preserving accurate style transfer. Quantitatively, the VLM leakage score drops from 2.674 to 0.522, confirming that the attention-level regularization effectively suppresses cross-reference semantic contamination. Frequency-aware RoPE modulation. Figure 9 compares models trained with and without RoPE modulation under the CRef+SRef setting. Without modulation, content-specific features from the style image (e.g., object silhouettes or texture patterns) bleed into the output via positional correspondence. Enabling the modulation suppresses this frequencydomain leakage while maintaining strong style transfer. The VLM leakage score decreases from 1.047 to 0.453, consistent with the visual comparison. Training data. To validate our data pipeline, we compare against a variant trained on OmniStyle data while keeping all other settings identical. Table 4 shows clear improvements on style-oriented metrics: ONEIG increases from 0.277 to 0.468, VLM-Style from 3.038 to 7.142, and Ver-S from 0.186 to 0.482. The qualitative comparison in Figure 10 reveals an even more pronounced perceptual difference. Across the displayed examples, the OmniStyle-trained variant often captures only part of the reference appearance, whereas our model more consistently reproduces the overall visual style while preserving the source content. This qualitative advantage is particularly evident for complex and long-tail styles.
8. Discussion and Conclusion Our results demonstrate that the core challenge of dualreference generation is not improving style strength or
Content
Style
Ours
Gemini-3
Seedream
Telestyle
USO
Figure 6. Qualitative Comparison on Style-reference (SRef) Generation. Our model achieves faithful stylistic alignment while avoiding the structural artifacts and semantic leakage observed in competing baselines.
content fidelity in isolation, but stably suppressing crossreference semantic contamination under multi-condition control. Experiments and attention analyses confirm that content leakage arises through distinct mechanisms in different settings: attention over-allocation in style-reference generation, and positional correspondence in dual-reference generation. This motivates our two-stage design where data construction and training constraints are co-optimized for each failure mode. Several limitations remain. First, community LoRA
quality follows a long-tailed distribution and evolves rapidly, making automated curation an ongoing challenge. Second, stylistic semantics across different base models still exhibit domain shift, limiting cross-model transferability. Third, existing evaluation metrics, including the proposed Verification Score, remain insufficient for fine-grained characterization of style-content conflict intensity. Future work will focus on automated LoRA quality assessment, cross-model style alignment, and more granular leakage metrics. In summary,
Gemini-3
Figure 7. Qualitative Comparison on Dual-reference (CRef+SRef) Generation. Given a content reference (col. 1) and a style reference (col. 2), our method (col. 3) faithfully preserves the layout of the CRef while transferring the artistic attributes of the SRef, exhibiting substantially less semantic leakage than all baselines.
Figure 8. Ablation on the Attention-map Enrichment Loss. Without the constraint (left of each pair), the model hallucinates content from the style reference. Enabling it (right) cleanly disentangles style from content.
Table 4. Dataset Ablation on the SRef Benchmark. Both models use the same architecture and training setup; only the data source differs. Metric DINOv2 ↑ CAS ↓ ONEIG ↑ CSD ↑ LAION-Aes ↑ V2.5-Aes ↑ VLM-Style ↑ VLM-Content ↑ Ver-S ↑ Ver-C ↑
OmniStyle Ours 0.897 0.762 0.277 0.577 5.957 5.527 3.038 9.669 0.186 0.972
0.809 1.175 0.468 0.639 6.302 5.709 7.142 8.919 0.482 0.928
we present FreeStyle, a complete framework for style- and content-dual-reference generation comprising a communityLoRA-based data pipeline, a two-stage training strategy with stage-specific disentanglement mechanisms, and a systematic benchmark. Extensive experiments show that jointly optimizing data, training constraints, and evaluation leads to more robust and balanced dual-reference generation than improving any single component alone.
9. Data Usage and Ethics Disclaimer This work is conducted strictly for non-commercial academic research purposes. Portions of the data used in this study, including reference images, community LoRA weights, and associated metadata, were collected from publicly accessible websites and open online repositories. These materials were used solely to study and evaluate reference-
Content
Style
Base
RoPE Module
Figure 9. Ablation on Frequency-aware RoPE Modulation. Without modulation (left of each pair), positional correspondence causes content leakage from the style reference. Enabling it (right) suppresses leakage while preserving style fidelity.
Figure 10. Dataset Ablation: the Proposed LoRA-based Pipeline vs. OmniStyle. Our model produces more faithful style transfer with richer stylistic detail, whereas the OmniStyle-trained variant shows weaker expressiveness, especially on complex and long-tail styles.
based image generation, and no part of this work is intended for, or has been deployed in, any commercial product or service. We do not claim ownership of any third-party content collected during data construction. All copyrights, trademarks, and other intellectual-property rights in the original images, models, and metadata remain with their respective owners. Where applicable, we have made reasonable efforts to respect the terms of use and licenses associated with the source materials, and we use such content only in the limited form necessary for scientific analysis, fair-use academic evaluation, and reproducibility of our reported results. Any names, styles, or trademarks that may appear are the property of their respective holders and are referenced only for identification and research discussion. The released benchmark, prompts, and code are provided “as is,” without any warranty of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, or non-infringement. The authors and their affiliated institutions accept no liability for
any direct, indirect, incidental, or consequential damages arising from the use, misuse, or inability to use the data, models, or methods described in this paper. Users who obtain or reproduce any part of this work are solely responsible for ensuring that their own use complies with all applicable laws, regulations, platform terms of service, and third-party rights in their respective jurisdictions. If any rights holder believes that specific content should not be included, we will promptly remove the corresponding material upon a reasonable request. By using the data, benchmark, or code associated with this paper, users acknowledge and agree to the terms of this disclaimer.
References [1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 8 [2] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 862–871, 2021. 3 [3] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. 5, 6, 9 [4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun 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. Qwen2.5vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6 [5] Black Forest Labs. Flux.2 [klein]: Towards interactive visual intelligence. https://blackforestlabs.ai, 2026. Official model announcement, January 15, 2026. 3 [6] 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 [7] Jingjing Chang, Yixiao Fang, Peng Xing, Shuhan Wu, Wei Cheng, Rui Wang, Xianfang Zeng, Gang Yu, and Hai-Bao Chen. Oneig-bench: Omni-dimensional nuanced evaluation for image generation. Advances in Neural Information Processing Systems, 38, 2026. 9 [8] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561– 26573, 2021. 3 [9] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2 [10] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting largescale diffusion models for style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8795–8805, 2024. 3 [11] Civitai. Civitai: A generative ai model-sharing platform. https://civitai.com/. Accessed: June 16, 2026. 4 [12] comfyanonymous. ComfyUI: A powerful and modular stable diffusion gui and backend. https://github.com/ comfyanonymous/ComfyUI, 2023. GitHub repository. 4 [13] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition, pages 11326–11336, 2022. 2 [14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021. 2 [15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 [16] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In International Conference on Learning Representations, 2017. 2 [17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024. 6 [18] Xiaoman Feng, Mingkun Lei, Yang Wang, Dingwen Fu, and Chi Zhang. Cleanstyle: Plug-and-play style conditioning purification for text-to-image stylization. arXiv preprint arXiv:2602.20721, 2026. 3 [19] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel CohenOr. Implicit style-content separation using b-lora. In European Conference on Computer Vision, pages 181–198. Springer, 2024. 3 [20] Zhoujie Fu, Xianfang Zeng, Jinghong Lan, Xinyao Liao, Cheng Chen, Junyi Chen, Jiacheng Wei, Wei Cheng, Shiyu Liu, Yunuo Chen, et al. imontage: Unified, versatile, highly dynamic many-to-many image generation. arXiv preprint arXiv:2511.20635, 2025. 3 [21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2 [22] Junyao Gao, Sibo Liu, Jiaxing Li, Yanan Sun, Yuanpeng Tu, Fei Shen, Weidong Zhang, Cairong Zhao, and Jun Zhang. Megastyle: Constructing diverse and scalable style dataset via consistent text-to-image style mapping. arXiv preprint arXiv:2604.08364, 2026. 3 [23] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414–2423, 2016. 2 [24] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 [25] Google DeepMind. Nano banana pro. https : / / deepmind . google / models / gemini - image / pro/, 2025. Official model page. 3 [26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 3
[27] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 3 [28] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 [29] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2 [30] Edward J Hu, Yelong Shen, Phil Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2 [31] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, pages 1501–1510, 2017. 2, 9 [32] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping selfattention. arXiv preprint arXiv:2402.12974, 2024. 3 [33] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2 [34] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019. 3 [35] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020. 3 [36] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051–10060, 2019. 3 [37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of textto-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931–1941, 2023. 2 [38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956–1981, 2020. 4 [39] Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 4 [40] LAION-AI. LAION-Aesthetics Predictor V1. https:// github.com/LAION- AI/aesthetic- predictor, 2022. GitHub repository, accessed June 16, 2026. 9 [41] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3809–3817, 2019. 3
[42] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. In NeurIPS, 2017. 2 [43] LiblibAI. Liblibai: An ai creation and model-sharing platform. https://www.liblib.art/. Accessed: June 16, 2026. 4 [44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 6 [45] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649–6658, 2021. 2 [46] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 6 [47] Fujun Luan, Sylvain Paris, Eli Shechtman, and Kavita Bala. Deep photo style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4990–4998, 2017. 3 [48] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3 [49] Aryan Mikaeili, Or Patashnik, Andrea Tagliasacchi, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Untwisting rope: Frequency control for shared attention in dits. arXiv preprint arXiv:2602.05013, 2026. 8, 10 [50] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2 [51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 [52] OpenAI. Gpt image 1.5. https://platform.openai. com/docs/models/gpt-image-1.5, 2025. Official model documentation. 3 [53] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 9 [54] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2 [55] Sang Hyun Park, Jun Young Koh, Junha Lee, Joy Song, Dongha Kim, Hoyeon Moon, Hyunju Lee, and Min Song. Illustrious: an open advanced illustration model. arXiv preprint arXiv:2409.19946, 2024. 4
[56] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. 6 [57] 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, pages 1862–1874, 2024. 2, 5 [58] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8693–8702, 2024. 3 [59] Qwen Team. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025. 3, 4, 10 [60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PmLR, 2021. 9 [61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2 [62] 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. 2 [63] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500–22510, 2023. 2 [64] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479–36494, 2022. 2 [65] Team Seedream, Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xiaoxia Hou, Weilin Huang, Yixuan Huang, et al. Seedream 4.0: Toward nextgeneration multimodal image generation. arXiv preprint arXiv:2509.20427, 2025. 10 [66] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422–438. Springer, 2024. 3 [67] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8242–8250, 2018. 2 [68] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn En-
tis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. In NeurIPS, 2023. 2 [69] Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024. 9 [70] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 [71] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 8 [72] Tensor.Art. Tensor.art: An ai model-sharing and generation platform. https://tensor.art/. Accessed: June 16, 2026. 4 [73] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. arXiv preprint arXiv:1603.03417, 2016. 2 [74] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6 [75] Verb. Aesthetic Predictor V2.5: A SigLIP-based aesthetic score predictor. https : / / github . com / discus0434/aesthetic-predictor-v2-5, 2024. Version 2024.12.18.1, GitHub repository. 4, 9 [76] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3 [77] Ye Wang, Ruiqi Liu, Jiang Lin, Fei Liu, Zili Yi, Yilin Wang, and Rui Ma. Omnistyle: Filtering high quality style transfer data at scale. arXiv preprint arXiv:2505.14028, 2025. 3 [78] Zihao Wang, Yuxiang Wei, Xinpeng Zhou, Tianyu Zhang, Tao Liang, Yalong Bai, Hongzhi Zhang, and Wangmeng Zuo. Premier: Personalized preference modulation with learnable user embedding in text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 29146–29156, 2026. 3 [79] Shaojin Wu, Mengqi Huang, Yufeng Cheng, Wenxu Wu, Jiahe Tian, Yiming Luo, Fei Ding, and Qian He. Uso: Unified style and subject-driven generation via disentangled and reward learning. arXiv preprint arXiv:2508.18966, 2025. 3 [80] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European conference on computer vision, pages 189–206. Springer, 2022. 3 [81] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Contentstyle composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 2, 3, 9 [82] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ipadapter: Text compatible image prompt adapter for text-toimage diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2
[83] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9036–9045, 2019. 3 [84] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836–3847, 2023. 2 [85] Shiwen Zhang, Xiaoyan Yang, Bojia Zi, Haibin Huang, Chi Zhang, and Xuelong Li. Telestyle: Content-preserving style transfer in images and videos. arXiv preprint arXiv:2601.20175, 2026. 3 [86] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8035–8045, 2022. 3 [87] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146–10156, 2023. 3 [88] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223–2232, 2017. 3 [89] Zhuofan Zong, Dongzhi Jiang, Bingqi Ma, Guanglu Song, Hao Shao, Dazhong Shen, Yu Liu, and Hongsheng Li. Easyref: Omni-generalized group image reference for diffusion models via multimodal llm. arXiv preprint arXiv:2412.09618, 2024. 3
A. Appendix A.1. Model Architecture without Extra Image Encoders Our style reference integration does not rely on an additional external image encoder. This design choice is motivated by the observation that existing image encoders possess limited capabilities in clustering and classifying artistic styles, and relying on them for style feature extraction might ultimately impair the model’s performance. Therefore, during training, we disentangle the model’s attention layers and apply our constraints directly on the first block. The overall architecture of our model is summarized in the middle panel of Figure 5. To further illustrate why we do not employ other external image encoders, we conducted a clustering experiment on stylistic features. As shown in Figure 12, we selected 4 different style LoRAs and combined their respective trigger words with the same set of 150 base prompts to generate test images via text-to-image synthesis. Since style LoRAs can be viewed as clustering centers for different stylistic concepts, we extracted the features of these generated images using several common image encoders (e.g., CLIP, DINOv2) as well as the VAE, and visualized them using t-SNE dimensionality reduction. The observation reveals that compared to other image encoders, the features extracted by the VAE can relatively clearly distinguish the clusters of different styles. This further validates the rationale behind our architectural choice to directly utilize the latent features extracted by the VAE for explicit guidance.
Block 0
Block 30
Block 40
Block 50
Block 59
Step 14
Step 21
Figure 11. Semantic Analysis of Intermediate Features across Transformer Blocks and Denoising Steps. Early blocks (block 0) primarily encode global semantics and layout, whereas later blocks focus on fine-grained details. As semantic leakage at block 0 cannot be corrected by subsequent layers, we impose our attention constraint at this stage. Table 5. Detailed Statistics of the Curated CRef and SRef Dataset Composition across Different Base Models. The total number of triplets is derived by multiplying the number of dual LoRA combinations, style LoRAs, and content LoRAs by their respective average number of generated images.The “Total Triplets” reported in the table are calculated based on the Cartesian product. However, the actual number of samples released is obtained by sufficiently sampling the style reference images rather than exhaustively enumerating all possible combinations.
A.2. Why Constrain the First Block To further motivate our design choice of applying the attention constraint on the first block, Figure 11 visualizes the intermediate feature representations across different denoising steps and block indices. It can be clearly observed that early blocks (especially the first block) determine the overall spatial layout and semantic composition of the generated image at the very beginning of the denoising process. As the block index increases, features progressively refine texture and local details, but the global structure has already been established in the early blocks. This observation indicates that if content leakage occurs at the first-block level, it will directly corrupt the semantic layout of the entire generated image, and subsequent blocks can hardly correct it. Therefore, applying the attention constraint on the first block is the most effective intervention point for suppressing crossreference semantic leakage.
Qwen FLUX Illustrious
This figure reports phrase-level word cloud of the LoRA trigger words for style and content, providing additional evidence for the vocabulary diversity used in data construction.
Block 20
Step 7
Base Model
A.3. Trigger-Word Statistics
Block 10
Step 0
Style LoRAs
Dual LoRA Combinations
Content LoRAs
Count Avg. Images Count Avg. Images Count Avg. Images 608 43,750 24,127
8.85 3.50 3.15
53 1,460 191
8.00 9.11 9.36
19 91 799
13.12 9.16 15.20
Total Triplets ∼935K ∼38.80M ∼68.39M
A.4. CRef and SRef Dataset Composition To elaborate on the scale of our constructed SRef and CRef dataset, Table 5 reports the detailed data composition across the three base models (Qwen, FLUX, and Illustrious). The total number of triplets is calculated as the product of the retained dual-reference LoRA combinations and the average number of generated images per combination, per style LoRA, and per content LoRA. For instance, in the Qwen setting, we retain 608 valid LoRA combinations (with an average of 8.85 images per combination). Combined with 53 style LoRAs (averaging 8.00 images) and 19 content LoRAs (averaging 13.12 images).This multiplicative compositional mechanism guarantees immense diversity at both the content and stylistic levels.
Figure 12. t-SNE Visualization of Style Clustering in Different Encoder Feature Spaces. Test images are synthesized from 4 distinct style LoRAs using the same 150 prompts, with each LoRA serving as a style cluster center. Compared with CLIP and DINOv2, VAE latent features yield the clearest separation of style clusters, motivating our encoder-free design that directly guides generation with VAE latents.
A.5. Additional Visual Comparisons for SRef
A.7. Extended CRef+SRef Comparisons (Group II)
To further corroborate the style-reference (SRef) comparison presented in the main text, Figure 14 provides additional qualitative examples under highly challenging conditions. These supplementary results cover a broader spectrum of artistic styles (e.g., oil painting, watercolor, 3D rendering, and abstract art) as well as more complex object structures. As demonstrated, even when confronted with long-tail style distributions, our model exhibits exceptional consistency in style transfer while consistently maintaining an extremely low rate of content leakage.
Figure 22 provides further examples spanning paper-cutting, japonism, fauvism, printmaking, flat vector, smooth clay, multilayer paper, leaf art, constructive illustration, and handmade clay styles. These long-tail artistic styles are particularly challenging, yet our method consistently outperforms competing baselines.
Gr ap h
d Moo
M ot
Co mp o
sit io
n
e
Content
Visual
r
ion ss pre Ex
Rendering
re &
Format
i on
Scene
i on al Media
u Text
Style
Tr ad it
Characte
ht
ics
r Lig
tic & T the on Aes e
lo Co
To complement the examples presented in Figure 1 in the main text, Figures 23–26 provide additional showcases of the style–content dual-reference triplet data generated through our LoRA combinations. These abundant samples further demonstrate the effectiveness of our data pipeline in producing high-quality triplets with exceptional stylistic diversity and accurate content preservation.
osp he r
Figure 15 presents additional dual-reference generation results covering styles such as 3D modeling, origami, pointillism, geometric abstraction, and graffiti. These examples further demonstrate our method’s strong generalization across diverse artistic domains while maintaining content fidelity.
A.9. Additional Triplet Data Showcases
At m
A.6. Extended CRef+SRef Comparisons (Group I)
A.8. System prompt in VLM-based metrics
Det
ail
Figure 13. Trigger-word Word Clouds in the Collected LoRA Metadata. The style (left) and content (right) word clouds together evidence the vocabulary diversity underlying our prompt-pool construction.
Content
Style
Ours
Gemini-3
Seedream
Telestyle
USO
Figure 14. Additional Qualitative Comparisons for Style-reference (SRef) Generation across Diverse Artistic Domains, including oil painting, watercolor, 3D rendering, and abstract art. Our method maintains consistent style transfer and low content leakage even under long-tail style distributions.
Gemini-3
Figure 15. Extended Dual-reference (CRef+SRef) Comparisons (Group I). Styles include 3D rendering, origami, pointillism, geometric abstraction, children’s drawing, graffiti, and quick sketch. Our method achieves faithful style transfer with minimal content leakage across all cases.
Content Similarity Evaluation Prompt
Prompt: [Role] You are an objective Image Content Adjudicator. Your task is to evaluate the subject-content consistency between two images based strictly on visible evidence. [Evaluation Criterion] Evaluate how consistent Image B is with Image A in terms of subject content and semantic theme, regardless of differences in visual style. Content consistency includes: - The identities and attributes of people or other subjects. - The categories, attributes, and counts of prominent objects. - The spatial arrangement and relative positions of major elements. - The background environment and overall scene category. - The actions, poses, gestures, and interactions depicted. - Meaningful text, logos, numbers, or symbols visible in the scene. Ignore stylistic differences, including rendering style, brushwork, lighting, color grading, image resolution, noise, and general aesthetics. [Task] Compare Image B with Image A and assign an integer content-similarity score from 0 to 10. Base the evaluation on the following observable aspects: 1. Human identity and attributes: - Facial identity, facial structure, and distinguishing features. - Hairstyle, hair color and length, body shape, and physical build. - Clothing categories, colors, patterns, logos, and accessories. 2. Objects and attributes: - Object categories, colors, materials, sizes, and counts. - The presence or absence of important props. - Legible text or logos associated with prominent objects.
3. Spatial layout and composition: - The relative positions of major subjects and objects. - Subject scale, viewpoint, and scene arrangement, allowing for reasonable compositional variation. 4. Background and scene category: - Indoor or outdoor setting. - Type of location and major structures, furniture, or landmarks. 5. Actions, poses, and interactions: - Human or animal poses, gestures, actions, and interactions. 6. Text, logos, and symbols: - Words, numbers, logos, or symbols that are important to the semantic meaning of the scene. Apply a strict scoring standard. Do not assign a high score unless the major subjects and semantic content are clearly and accurately consistent. If subject identity or important content cannot be verified because of blur, occlusion, or insufficient visual evidence, assign a conservative score. [Scoring Rubric - 11 Levels] - 0: The two images are completely unrelated. - 1-3: The images are mostly inconsistent and share only minimal content. - 4-6: The images are partially consistent but contain significant content mismatches. - 7-9: The images are mostly consistent, with only minor content differences. - 10: All major subjects, attributes, actions, objects, and scene elements are fully consistent. [Output Format] Return ONLY one line in the following format: <score>@<reason> The score must be an integer from 0 to 10. The reason must contain one or two short sentences describing specific and directly observable similarities or mismatches.
Figure 16. Prompt template for evaluating subject-content consistency between a reference image and a generated image.
Style Similarity Evaluation Prompt
Prompt: [Role] You are an objective Image Style Adjudicator. Your task is to evaluate the overall visual-style consistency between two images based strictly on observable stylistic evidence. [Evaluation Criterion] You will be given two images, Image A and Image B. Evaluate how consistent their visual styles are. Assess only visual style. Do not consider whether the images depict the same people, objects, actions, backgrounds, or scenes. The semantic content of the two images may be completely different. The evaluation must jointly consider the following five dimensions: 1. BRUSHSTROKE (brushwork, line quality, and edge treatment) - Line thickness, line stability or roughness, outlining methods, and the visibility of brush marks. - Hard or soft edges, the presence of contours, and characteristic techniques such as smearing, splashing, pencil, ink, or pen strokes. 2. TEXTURE (surface texture, material rendering, grain, noise, and canvas or paper quality) - Fine-grained surface characteristics, such as sandpaper-like grain, film noise, watercolor-paper fibers, or oil-canvas texture. - The representation of fine material structures, such as layered paint, spray dots, halftone patterns, or embossed surfaces. - Texture does not refer to color schemes or object contours. 3. COLOR (color palette, color distribution, temperature, saturation, and contrast) - Dominant hues, warm or cool color tendencies, saturation level, and the strength of light-dark contrast. - Similarity in color distribution, including the proportions and coverage areas of dominant colors, large background regions, and major color blocks. - Color does not refer to surface grain, texture, or line quality. 4. SHAPE (shape language and form-construction conventions) - The degree of geometric abstraction or realism, exaggerated proportions, and the use of sharp or rounded contours.
- Structural simplification conventions, such as flat cartoon, chibi, minimalist geometric, or realistic construction. 5. PATTERN (recurring motifs and decorative organization) - Repeated patterns, decorative elements, and recurring symbolic motifs, such as fixed ornaments, repeated decorative lines, or patterned backgrounds. - The density, repetition rules, and organization of decorative elements. Do not evaluate aesthetic quality, artistic merit, or whether either image is visually appealing. [Task] Compare Image B with Image A and assign one integer score from 0 to 10 representing their overall visual-style similarity. The final score must reflect a holistic assessment across all five dimensions rather than any single dimension in isolation. Apply a strict scoring standard. Do not assign a high score unless the key stylistic characteristics are clearly and accurately aligned. [Scoring Rubric - 11 Levels] - 0: The visual styles are completely inconsistent, with essentially no meaningful stylistic correspondence. - 1-3: The styles are mostly inconsistent and share only weak or incidental similarities. - 4-6: The styles are partially similar but contain clear omissions or major differences in important stylistic characteristics. - 7-9: The styles are largely consistent, with only a small number of minor differences. - 10: The visual styles are fully consistent, with the key stylistic characteristics closely matching across all dimensions. [Output Format] Return ONLY one line in the following format: <score>@<reason> The score must be an integer from 0 to 10. The reason must consist of one or two short sentences identifying specific and directly observable stylistic evidence. Do not output any additional text, labels, punctuation, or line breaks.
Figure 17. Prompt for Evaluating Overall Visual-style Similarity between a Reference Image and a Generated Image.
Content Verification Score Prompt
Prompt: [Role] You are a strict Image Content Adjudicator who evaluates only subject content and semantic theme. You must completely ignore differences in visual style, including artistic style, linework, color treatment, rendering method, resolution, filters, and other stylistic properties. [Evaluation Criterion] Determine whether Image A and Image B are highly consistent in terms of the specific subjects, objects, actions, scenes, and semantic themes depicted. Focus on what is shown and what is happening, rather than how the image is rendered. Apply the following criteria: 1. Human-centered images - Determine whether the images depict the same character or two extremely similar characters. - Consider observable attributes such as gender presentation, age group, body build, hairstyle, hair color, skin tone, clothing type, dominant clothing colors, and major accessories. - Moderate differences in pose, orientation, or camera viewpoint are acceptable. - If the images clearly depict different people or substantially different character designs, they are inconsistent. 2. Single-object images - Determine whether the main objects have the same specific category, shape, and structural configuration. - For example, both objects may be sports cars, SUVs, round tables, or structurally similar buildings. - Differences in color are acceptable when the object category and overall form remain highly similar. - Merely belonging to a broad category, such as both being vehicles, houses, or cups, is insufficient when their types or structures are clearly different. 3. Complex scenes - Consider the scene category, combination of major elements, spatial
layout, and central semantic theme. - For example, two images may be considered consistent if both depict a person standing in the center of a nighttime city street, with tall buildings and neon signs in the background. - Merely sharing a broad indoor or outdoor setting is insufficient when the central subjects, composition, and major objects are substantially different. 4. Independence from visual style - Visual style must not affect the decision. - For example, a realistic photograph and an anime or cartoon illustration must be judged consistent when their subject content and semantic theme are highly aligned. - Never classify a pair as inconsistent solely because their visual styles differ. [Task] Compare Image A and Image B and make a binary decision regarding their subject-content and thematic consistency. Output 1 when the images depict the same character, the same specific type of object with a highly similar structure, or the same specific scene and semantic theme. Output 0 when the images share only a broad category, such as both containing a person or a vehicle, but their primary subjects or semantic content are clearly different. [Decision Rubric - 2 Levels] - 0 (Inconsistent): The primary subjects, objects, scene structure, actions, or semantic themes are not highly consistent. - 1 (Consistent): The primary subjects and semantic themes are highly consistent, regardless of differences in visual style. [Output Format] Return ONLY one character: 0 or 1 Do not output any explanation, label, space, line break, punctuation, or JSON.
Figure 18. Prompt Template for Content verification Consistency between Two Images..
Style Verification Score Prompt
Prompt: [Role] You are an experienced Image Style Adjudicator who evaluates only visual style and modes of visual representation. Your assessment must focus on stylistic properties, including the perceived medium, rendering technique, material appearance, linework and brushwork, color treatment, lighting and contrast, post-processing, image noise and grain, and methods of detail representation. You must ignore the identities of people or objects, the meanings of actions, narrative semantics, scene categories, and whether the depicted content or composition is similar. [Evaluation Criterion] Determine whether Image A and Image B belong to the same stable visual style or style family. Use a permissive consistency standard: the images may still be judged stylistically consistent when their core style mechanisms remain aligned, even if their subjects, scenes, compositions, viewpoints, or levels of detail differ. The following differences are acceptable and should not independently cause an inconsistent judgment: - Different subjects, objects, or scenes. - Different compositions, viewpoints, camera positions, or crop ranges. - Minor hue shifts, brightness changes, contrast variations, or local color-grading differences. - Different levels of detail, resolution, cropping, mild compression, noise, or grain. Judge the images as inconsistent only when there is a clear change in the underlying style mechanism, such as: - Realistic photography versus illustration or computer rendering. - A major change in the linework system, such as outlined versus lineless rendering, thick versus thin outlines, or comic-style lines versus watercolor edges. - A major change in material or texture generation, such as impasto oil painting, flat cel shading, glossy three-dimensional rendering, pixel art, or pointillism. - A major change in the lighting model, such as hard cinematic lighting, soft diffuse illustrative lighting, or high-contrast neon lighting. - A fundamental change in color strategy, such as muted vintage colors,
highly saturated candy colors, or monochrome sketch rendering. [Task] Compare Image A and Image B only in terms of visual style and determine whether they belong to the same style family. Base the judgment primarily on the following dimensions: 1. MEDIUM AND RENDERING METHOD - Photography, three-dimensional rendering, digital illustration, watercolor, oil painting, impasto, cel shading, pixel art, sketch, or other visual media and rendering paradigms. 2. BRUSHWORK AND LINE SYSTEM - Presence or absence of outlines. - Line thickness, stability or roughness, edge treatment, visible brush marks, and stroke granularity. 3. MATERIAL AND TEXTURE GENERATION - Surface appearance, texture construction, image noise, grain, and the organization of fine visual details. 4. LIGHTING MODEL AND CONTRAST - Hard or soft shadows, diffuse or specular reflection, volumetric lighting, contrast level, and the general method of modeling light. 5. COLOR STRATEGY - Saturation, hue preferences, overall tonal balance, color temperature, and grading conventions such as vintage, warm, cool, or neon treatments. Composition and viewpoint are secondary dimensions. Differences in camera position, framing, perspective, or crop should not directly produce an inconsistent judgment. [Decision Rubric - 2 Levels] - 0 (Inconsistent): The images exhibit a clear change in one or more core style mechanisms, such as their medium, rendering paradigm, linework system, texture-generation method, lighting model, or overall color strategy. - 1 (Consistent): Most major style dimensions are aligned and the images belong to the same style family, even if their subjects, scenes, compositions, viewpoints, or detail densities differ. [Output Format] Return ONLY one character: 0 or 1 Do not output any explanation, label, space, line break, punctuation, or JSON.
Figure 19. Prompt Template for Style verification Consistency between Two Images.
Instruction-Following Evaluation Prompt
Prompt: [Role] You are an objective Image Editing Adjudicator. Your task is to evaluate how accurately a final edited image follows a given editing instruction, based strictly on visible evidence. [Evaluation Criterion] You will be given: 1. An image representing the final edited result. 2. A textual editing instruction. Evaluate the extent to which the final image satisfies the editing instruction. Award partial credit when only some parts of the instruction are correctly implemented. Do not assume that a requested modification has been completed unless it is clearly supported by visible evidence in the final image. Ignore differences in rendering style, brushwork, resolution, image noise, aesthetic quality, and overall visual appeal only when these properties are not explicitly requested by the instruction. If the instruction explicitly specifies visual properties such as color, lighting, mood, artistic style, camera lens, or texture, these properties must be treated as evaluation requirements. [Task] Assign a score from 0 to 10 indicating how well the final image follows the editing instruction. Internally perform the following evaluation procedure: 1. Decompose the instruction into atomic and visually verifiable requirements, including: - Main subjects and objects. - Key attributes, such as color, material, number, size, or identity. - Required actions and interactions. - Spatial and layout constraints, such as left, right, foreground, background, or relative position. - Required text, logos, numbers, or symbols, for which exact wording and appearance may be important. 2. Classify each requirement as either: - MUST-HAVE: an essential component of the requested edit. - NICE-TO-HAVE: a secondary or minor component. 3. Judge each requirement as: - Satisfied. - Partially satisfied. - Not satisfied.
- Contradicted. 4. Determine the final score by starting from 10 and applying penalties according to the severity of the errors: - Missing a MUST-HAVE requirement: subtract 2 to 4 points. - Partially satisfying a MUST-HAVE requirement: subtract 1 to 2 points. - Incorrect minor or NICE-TO-HAVE detail: subtract 0.5 to 1 point. - Directly contradicting the instruction: subtract 3 to 5 points. - Missing or incorrect required text or logo: subtract 3 to 6 points. Clamp the final score to the range from 0 to 10. Do not output the requirement decomposition, intermediate judgments, or penalty calculations. [Scoring Rubric] - 0-0.5: The final image is completely unrelated to, or directly opposed to, the editing instruction. Assign 0 only for a genuine total failure. - 1-3.5: The image has a slight relationship to the instruction, but most essential requirements are missing, incorrect, or contradicted. - 4-6.5: Some important parts of the instruction are implemented, but major omissions, incorrect edits, or misplaced elements remain. - 7-8.5: Most essential requirements are satisfied, with one or two notable errors or several minor issues. - 9-10: All essential requirements are satisfied, with at most negligible issues. Assign 10 only when no important requirement is missing or incorrect. [Output Format] Return ONLY one line in the following format: <score>@<reason> The score must be between 0 and 10 and may use increments of 0.5, such as 6.5, 7.0, or 8.5. The reason must consist of one or two short sentences identifying specific and directly observable satisfied, missing, incorrect, or contradicted requirements. Do not output any additional text, headings, intermediate steps, or line breaks.
Figure 20. Prompt for Accuracy Evaluation on Editing Instruction Following.
Content Leakage Prompt
Prompt: [Role] You are an objective Image Content Leakage Adjudicator. Your task is to evaluate the extent to which Image B reuses specific visual content from Image A, based strictly on observable evidence. [Evaluation Criterion] You will be given two images, Image A and Image B. Evaluate the degree of content leakage from Image A to Image B. Content leakage refers to whether Image B reuses specific subjects, character identities, object structures, scene layouts, poses, actions, camera compositions, or distinctive local elements from Image A, rather than merely adopting its visual style. Carefully distinguish between visual style and image content: - Acceptable stylistic similarity includes brushwork, material appearance, color systems, rendering methods, line conventions, and overall aesthetic character. - Content similarity includes the same character, hairstyle and clothing combination, specific object structure, scene layout, pose, composition, or combination of background elements. These properties must not be treated as purely stylistic similarity. [Task] Compare Image B with Image A and assign an integer score from 0 to 10 representing the degree of content leakage. Focus on the following observable leakage signals: 1. SUBJECT IDENTITY - Determine whether the images depict the same person, character, animal, or specific object. - For human subjects, consider gender presentation, hairstyle, facial features, clothing design, accessories, and body shape. 2. POSE AND ACTION - Determine whether Image B preserves the same or a highly similar pose, gesture, orientation, action, or interaction. 3. COMPOSITION AND VIEWPOINT - Consider whether the subject position, framing distance, camera angle, viewpoint, or cropping strategy is visibly reused. 4. SCENE AND OBJECT ARRANGEMENT - Consider whether key background objects, spatial layouts, foreground-background relationships, or combinations of decorative elements are substantially similar. 5. DISTINCTIVE DETAILS - Determine whether Image B reuses uncommon and specific combinations
of details, such as distinctive hair accessories, weapons, patterns, furniture structures, signs, or small background objects. Apply the following strict evaluation rules: - Do not assign a high score merely because the two images have similar visual styles. - Increase the score only when specific content from Image A is visibly carried over into Image B. - If the images share only a broad subject category, such as both being portraits of women, houses, or street scenes, but their specific content differs, assign a low score. - If Image B appears to adopt only the style of Image A while changing the subject, composition, and scene, assign a low score. [Scoring Rubric] - 0: No visible content leakage. The images may share a style, but their specific subjects, compositions, and scenes are different. - 1-2: Very weak content overlap. Similarities are primarily limited to a broad subject category or a few generic elements. - 3-4: Some visible content borrowing is present, but the main subjects, compositions, or scenes remain clearly different. - 5-6: Moderate content leakage. Multiple important content elements are similar, indicating that Image B visibly references specific content from Image A. - 7-8: Strong content leakage. The core subject, pose, composition, or scene organization is substantially reused, with only partial modifications. - 9-10: Extremely strong content leakage. Image B closely reproduces the specific content of Image A or introduces only minor modifications. [Output Format] Return ONLY one line in the following format: <score>@<reason> The score must be an integer from 0 to 10. The reason must consist of one or two short sentences identifying the specific content elements responsible for the score. Do not justify the score solely by stating that the visual styles are similar or different. Do not output any additional text, headings, punctuation, or line breaks.
Figure 21. Prompt for Discrete Content Leakage between Two Images.
Gemini-3
Figure 22. Extended Dual-reference (CRef+SRef) Comparisons (Group II). Styles include paper-cutting, japonism, fauvism, printmaking, flat vector, smooth clay, multilayer paper, leaf art, constructive illustration, and handmade clay. Our method demonstrates robust generalization to these challenging long-tail styles.
CONTENT
STYLE
TARGET
CONTENT
STYLE
TARGET
Figure 23. Additional Showcases of the Style–content Dual-reference Triplet Data (Part I). These samples offer a more comprehensive view of the high-quality and stylistically diverse dataset generated via our LoRA-combination pipeline.
CONTENT
STYLE
TARGET
CONTENT
STYLE
TARGET
Figure 24. Additional Showcases of the Style–Content Dual-reference Triplet Data (Part II).
CONTENT
STYLE
TARGET
CONTENT
STYLE
TARGET
Figure 25. Additional Showcases of the Style–Content Dual-reference Triplet Data (Part III).
CONTENT
STYLE
TARGET
CONTENT
STYLE
TARGET
Figure 26. Additional Showcases of the Style–Content Dual-reference Triplet Data (Part IV).