acceptodds
Under review as a conference paper at ICLR 2027

Chameleon: Style-Content Disentangled Framework for Cross-Domain Object Compositing

17 views

Abstract

Image compositing aims to seamlessly insert a foreground object into a background image, and recent advances in diffusion models have significantly enhanced the quality, especially when the foreground and background images come from the same domain (e.g., natural images). However, cross-domain compositing, where the foreground and background come from different domains, is relatively underexplored and remains challenging because the model must preserve the foreground object's identity while stylizing it to match the background domain. Existing cross-domain compositing approaches largely rely on training-free blending and refinement strategies. This is partly due to the lack of large-scale paired datasets for cross-domain compositing, limiting the development of training-based solutions. As a result, they are limited to tone-level alignment and often produce style-inconsistent or overstylized results. To overcome such limitations, we construct ChameleonDataset, the first large-scale training dataset for cross-domain compositing, with a comprehensive evaluation benchmark, built through a scalable data construction pipeline. Building on this, we propose Chameleon, a novel two-stage training-based cross-domain compositing framework. In the first stage, we propose Signed Style-Content Graph Learning (SSCGL) to train ChameleonEncoder, which effectively disentangles style and content representations. In the second stage, we introduce Style Gap-Aware Anisotropic Modulation (SGAM) into a diffusion transformer for effective cross-domain compositing, adaptively regulating how style tokens from the first-stage ChameleonEncoder are injected according to the foreground-background style gap, spatial region, and diffusion timestep. Our method outperforms state-of-the-art in-domain and cross-domain compositing models, sequential pipelines and commercial models, achieving improvements in both compositional plausibility and stylistic fidelity.

Primary area:
generative models
Keywords:
image compositing, diffusion models, cross-domain, image generation, style-content disentanglement, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jNiZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Chameleon, a training-based framework for cross-domain object compositing. It addresses the scarcity of training data by proposing a reverse data generation pipeline (ChameleonDataset) that uses real stylized composites and input degradation. To preserve foreground identity while adapting to background styles, the framework utilizes a style-content disentangled encoder trained via Signed Style-Content Graph Learning (SSCGL) and injects these representations into a diffusion transformer using Style Gap-Aware Anisotropic Modulation (SGAM).

Critical strengths
  1. Novel Reverse Data Pipeline: Inverting the standard synthesis process by starting from real-image composites and degrading inputs avoids training models on synthetic distribution artifacts. This data construction approach is well-motivated and supported by distribution error measurements.

  2. Effective Parameter-Efficient Architecture: The combination of SSCGL for disentangled feature representation and SGAM for conditional injection achieves strong qualitative and quantitative improvements while training a very small fraction of the network parameters via LoRA.

Critical weaknesses
  1. Reliance on an Unvalidated VLM Rubric: The claim of state-of-the-art performance relies heavily on an unweighted VLM-based scoring rubric. Since the test cases overlap with the user study, reporting the statistical correlation between the VLM rubric and human evaluations is necessary to validate this automated metric.

  2. Missing Clarification on Benchmark Separation: It is unclear whether the evaluation benchmark is strictly disjoint from the training data, given that both originate from the same underlying source corpora.

Questions for authors
  1. Dataset Disjointness: Is ChameleonDataset_ev strictly disjoint from the training split, and what are the exact steps taken to prevent data leakage between them?

  2. VLM Validation: What is the correlation coefficient between the VLM rubric scores and the human user study results across the overlapping evaluation cases?

4 comments
AuthorsOpenReview ↗

Response to Reviewer jNiZ (Part 1/2)

We appreciate Reviewer jNiZ's recognition of our reverse data pipeline and parameter-efficient framework, as well as their helpful comments.

[W1, Q2] VLM rubric validation against human preferences.

We thank the reviewer for this valuable suggestion. To validate the VLM-based evaluation rubric, we conducted a statistical correlation analysis between VLM rubric scores and human evaluations from our existing user study, comprising 50 participants and 30 overlapping evaluation cases (Tab. 4). Furthermore, to assess robustness across VLM judges, we additionally evaluate four VLMs: GPT-4.1 [9], GPT-5.5 [10], Opus 5.5 [11], and Gemini 3.1 Pro [12].

Specifically, the VLM scores each method on identity, style, and composition (0–3 per criterion, see Appendix H.1), while in the user study, participants select the best among five cross-domain compositing methods for each criterion, yielding human preference rates (win rates, see Appendix C.4). For each case, we compute Spearman's rank correlation (ρ\rho) [1] between the method rankings based on VLM scores and human preference rates, and average the correlations across all 30 cases. The user study's “Overall” criterion corresponds to the VLM's “Composition and Harmonization” criterion (Appendix C.4, L1992–1995, and Appendix H.1), while the total comparison uses the sum of the three criterion scores (identity, style, and composition) for each VLM judge and the mean of the corresponding human preference rates (identity, style, and overall).

VLM JudgeIdentity ρ\rho ↑Style ρ\rho ↑Comp. ρ\rho ↑Total ρ\rho ↑Tie (%)Pairwise (%) ↑
GPT-4o0.690.560.620.752383
GPT-4.10.720.500.600.821983
GPT-5.50.740.540.620.681374
Opus 5.50.750.650.630.741379
Gemini 3.1 Pro0.710.560.540.662271

The GPT-4o judge used for the VLM evaluation reported in our paper achieves a total-score correlation of ρ=0.75\rho = 0.75, with 83% pairwise agreement. To assess robustness across VLM judges, we further evaluate GPT-4.1, GPT-5.5, Opus 5.5, and Gemini 3.1 Pro using the same inputs, evaluation prompt, and scoring criteria. Across all five judges, total-score correlations range from 0.66 to 0.82, with 71%–83% pairwise agreement (Tab. 15). Tie rate measures the proportion of method pairs with identical VLM total scores, while pairwise agreement excludes pairs tied in either evaluation. Moreover, our method consistently achieves the highest mean total score across all five VLM judges (Tab. 16).

These results demonstrate consistent positive correlations between the VLM rubric and human preferences. We have added the validation protocol and results to Appendix C.4 and referenced this analysis in Sec. 4.1 of the revised manuscript (marked in red).

AuthorsOpenReview ↗

Response to Reviewer jNiZ (Part 2/2)

[W2, Q1] Dataset disjointness and prevention of data leakage.

We thank the reviewer for raising this important concern. **ChameleonDataset$_{ev}$ is constructed independently of ChameleonDataset$_{tr}∗∗,usingdistinctsourcecollections.Specifically,ChameleonDataset**, using distinct source collections. Specifically, ChameleonDataset_{tr}$ is built from existing stylization datasets ([2–4], Sec. 3.1, L211–213), whereas ChameleonDataset$_{ev}$ is independently collected from Pexels, Pixabay, and Unsplash ([5–7], Sec. 4.1, L455–457, and Appendix A.5). Thus, the two datasets do not originate from the same underlying source corpora.

To further verify disjointness and check for potential data leakage, we conduct multi-level near-duplicate detection using feature-level (SSCD, DINOv3) and pixel-level (image hashes, ORB) similarity checks (Fig. 16(a)). For each of the 156 evaluation images, we compute its highest SSCD similarity against the training set. Only four exceed the 0.45 threshold used for relative deduplication in DINOv2 [8] (Fig. 16(b)). We manually inspect threshold-flagged and top-ranked candidate pairs retrieved by all detectors, covering both foreground and background images, and find no confirmed near-duplicates (Fig. 16(c)).

These results support the separation of the evaluation benchmark from the training data. We have clarified the distinct data sources in Sec. 4.1 and added the verification protocol and results to Appendix A.5 (Fig. 16) of the revised manuscript (marked in red).

References.

[1] Spearman. The Proof and Measurement of Association Between Two Things. The American Journal of Psychology, 1987.

[2] Li et al. StyleTokenizer: Defining Image Style by a Single Instance for Controlling Diffusion Models. ECCV 2024.

[3] Liao et al. The ArtBench Dataset: Benchmarking Generative Models with Artworks. arXiv 2022.

[4] Ju et al. Human-Art: A Versatile Human-Centric Dataset Bridging Natural and Artificial Scenes. CVPR 2023.

[5] Pexels. Pexels. https://www.pexels.com, 2026.

[6] Pixabay. Pixabay. https://pixabay.com, 2026.

[7] Unsplash. Unsplash. https://unsplash.com, 2026.

[8] Oquab et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv 2023.

[9] OpenAI. Introducing GPT-4.1 in the API. OpenAI, 2025.

[10] OpenAI. Introducing GPT-5.5. OpenAI, 2026.

[11] Anthropic. Introducing Claude Opus 5.5. Anthropic, 2026.

[12] Google DeepMind. Gemini 3.1 Pro Model Card. Google DeepMind, 2026.

Thank you for the detailed response and the additional experiments. Most of my concerns have been addressed, so I'm inclined to maintain my score.

We sincerely thank the reviewer for taking the time to review our response and additional analyses. We are glad that we have addressed most of your concerns, and we appreciate your constructive feedback, which has helped improve our manuscript.

Reviewer_AcdLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Chameleon, a framework for cross-domain object compositing. First, the authors construct ChameleonDataset, which uses real images as ground-truth composites and generates the corresponding input conditions. The proposed method adopts a two-stage training framework. The first stage learns disentangled style and content representations, while the second stage uses these representations to guide image generation, adaptively modulating style injection based on the style gap, spatial region, and diffusion timestep. Experimental results demonstrate that the proposed method can insert foreground objects into stylized backgrounds while achieving relatively harmonious and style-consistent compositions.

Critical strengths
  • I like the reverse data construction pipeline, which uses real images as ground-truth composites rather than synthetic outputs. This provides more faithful supervision and encourages the model to learn natural and harmonious foreground-background integration.
  • The paper is well-written and easy to follow, with relatively comprehensive experiments to support the proposed method.
Critical weaknesses
  • I do not deny the effectiveness of Signed Style-Content Graph Learning (SSCGL), but it is inaccurate to claim that contrastive learning is unsuitable for this task. Positive pairs in contrastive learning do not necessarily have to be constructed through standard data augmentation; they can also be constructed using the Style/Content annotations adopted by the authors. Therefore, the discussion in lines 298–306 may not be sufficiently rigorous. Moreover, a contrastive loss carefully adapted to this task might perform no worse than the SSCGL loss.

  • I believe SSCGL can effectively disentangle style and content representations. However, style-content disentanglement is a well-established topic in style transfer, so disentangling these representations is not particularly challenging, especially with the improved representation capabilities of DINOv3.

  • It would be helpful to specify the generative backbone architectures used by the compared methods, as it is unclear whether any of them also fine-tune the weights of FLUX.1-dev.

  • Figure 1 highlights several advantages of the proposed method over existing approaches, including improved geometry alignment, shadow generation, reflections, and occlusion handling. However, while the paper provides explanations and supporting experiments for identity preservation, style consistency, and pose adaptation, the improvements in shadows, reflections, and occlusion handling are mainly demonstrated through qualitative examples. It remains unclear which components of the proposed framework are responsible for these capabilities. More targeted analysis and evaluation would help substantiate these claims.

Questions for authors
  • It would be helpful if the authors could more clearly explain the advantages of the SSCGL loss over conventional contrastive losses.
  • For other aspects, please refer to "Critical Weaknesses".
7 comments
AuthorsOpenReview ↗

Response to Reviewer AcdL (Part 1/3)

We sincerely thank Reviewer AcdL for the careful review and constructive feedback. We address the concerns through additional experiments, analyses, and clarifications, and summarize the corresponding changes in the revised manuscript (marked in red).

[W1, Q1] SSCGL vs. Contrastive Learning.

We thank the reviewer for highlighting that contrastive learning can also leverage style/content annotations to construct positive and negative pairs. In response, we have expanded the discussion in Sec. 3.3 (L308–316, marked in red) to include label-based contrastive learning and conducted additional experiments comparing SSCGL with SupCon [1] and label-based HCL [2] (Tab. 11).

Experimental comparison. To directly address this concern, we adapted SupCon and HCL to use the same style/content annotations as SSCGL. All methods share the frozen DINOv3 backbone, projection-head architectures, training data, batch construction, and training schedule, differing only in their training objectives. Following the dual-anchor evaluation protocol in Tab. 17, we evaluate the learned representations on newly sampled test splits (400 samples each). Detailed results are reported in Tab. 11.

MethodStyle Δ ↑Content Δ ↑
Frozen DINOv3−0.033+0.253
SupCon [1]+0.842+0.989
HCL (Label-based) [2]+0.801+0.894
SSCGL (Ours)+0.935+0.998

The results confirm that annotation-based contrastive learning is indeed effective, consistent with the reviewer's suggestion. Nevertheless, SSCGL achieves the highest disentanglement margins in both spaces, with clearer separation of style clusters in the embedding visualizations (Fig. 5).

Why SSCGL? SSCGL offers two key advantages for style-content disentanglement:

  1. Explicit orthogonality objective. While InfoNCE-based losses optimize relative positive-negative similarities, SSCGL directly aligns positive pairs and penalizes squared negative-pair similarity, explicitly encouraging negative-pair orthogonality (Sec. 3.3, Eq. 1). Appendix B.2 provides a theoretical characterization of SSCGL's factor-selective optimum, while Appendix B.3 directly compares its optimization behavior with InfoNCE (Figs. 17–18), complemented by empirical comparisons with label-based SupCon and HCL (Tab. 11).

  2. Factor-specific edge weighting. Beyond defining positive and negative pairs using annotations, SSCGL assigns factor-specific edge weights within the signed graph using distinct style and content similarities. Compared with uniform weighting on the same signed graph, this improves disentanglement margins in both style and content spaces (Tab. 20).

AuthorsOpenReview ↗

Response to Reviewer AcdL (Part 2/3)

[W2] Difficulty of style-content disentanglement with DINOv3.

As the reviewer notes, style-content disentanglement is a well-established topic in style transfer. However, content leakage has not been fully resolved, as recent works continue to address this issue [3, 4]. This challenge is particularly critical in cross-domain compositing, where background style must be transferred while preserving foreground identity without introducing background content (Fig. 3).

Existing style-transfer representations. To assess whether this challenge persists, we evaluate representations from six existing style-transfer methods using the dual-anchor disentanglement protocol (Tab. 21(a)). The margin measures factor-specific separation by comparing positive pairs sharing the target factor (style or content) with hard negatives sharing the confounding factor.

MethodStyle Δ\Delta ↑Content Δ\Delta ↑
StyleSSP−0.231+0.229
MaskST−0.299+0.297
CSGO+0.028−0.028
StyleShot+0.148−0.171
USO+0.134−0.105
StyleExpert+0.223−0.195
Ours+0.926+0.996

Existing style-transfer methods show disentanglement in either style or content, but none of the six achieves positive disentanglement margins on both splits, indicating that existing style-oriented representations do not consistently separate style and content.

Backbone generalization. To examine whether SSCGL's effectiveness primarily stems from DINOv3's strong pretrained representations, we extend SSCGL to four additional frozen backbones from the CLIP and SigLIP families (Tab. 21(b)).

BackboneStyle Δ\Delta ↑Content Δ\Delta ↑
DINOv3 ViT-L/16 (Frozen)−0.030+0.249
+ SSCGL+0.926+0.996
CLIP ViT-L/14 (Frozen)−0.093+0.215
+ SSCGL+0.909+0.995
CLIP ViT-H/14 (Frozen)−0.201+0.200
+ SSCGL+0.920+0.997
SigLIP so400m (Frozen)−0.200−0.165
+ SSCGL+0.812+0.996
SigLIP2-L/16-256 (Frozen)−0.155+0.153
+ SSCGL+0.868+0.996

Notably, all five frozen backbones exhibit negative style disentanglement margins, including DINOv3 (−0.030). After applying SSCGL, all five achieve positive margins in both spaces. These results demonstrate that SSCGL's disentanglement effectiveness extends across different backbone families rather than being attributable solely to DINOv3.

We have added these analyses to Appendix D.1 (Tab. 21) and included a corresponding reference in Sec. 3.3 (L298–300, marked in red).

[W3] Backbones of compared methods.

We thank the reviewer for suggesting a clearer specification of the generative backbones and adaptation strategies. We have added a Backbone column to Tab. 2 of the revised manuscript (marked in red), reporting each method's backbone and training strategy as used in its original implementation. Specifically, DreamFuse also uses FLUX.1-dev with LoRA fine-tuning, while SHINE uses it in a training-free setting.

AuthorsOpenReview ↗

Response to Reviewer AcdL (Part 3/3)

[W4] Shadows, reflections, and occlusion handling.

We clarify two complementary aspects of our framework that support these capabilities: real-image supervision from our reverse data construction pipeline and a positional affine transformation that allows compositional effects to extend beyond the foreground silhouette.

Real-image supervision. As described in Sec. 3.1 (L211–221, Fig. 2), our reverse pipeline uses real composites IcI_c as supervision targets. As illustrated in Fig. 13(a), ObjectClear removes foreground objects and their associated shadows and reflections from Ib′I_b' in the shown examples, while the real-image target IcI_c retains these effects. This input–target asymmetry provides direct denoising supervision for their reconstruction. For occlusion, the foreground reference If′I_f' depicts the object without surrounding occluders, whereas IcI_c preserves its partial occlusion by scene elements, encouraging the model to learn foreground-background occlusion relationships.

Positional affine transformation. As described in Sec. 3.4 (L398–403) and Appendix E, our method warps foreground latent positional indices to specify the target placement rather than directly blending foreground pixels. Unlike mask-constrained blending, this transformation does not restrict generation to the foreground silhouette. Fig. 13(b,c) illustrates this distinction from AIComposer's mask-constrained pixel-space blending.

Targeted quantitative evaluation. We evaluate shadow and reflection fidelity on DESOBAv2 (SGDiffusion [5]) and DEROBA (RGDiffusion [6]) using Local RMSE (LR↓) and Local SSIM (LS↑), and occlusion preservation on CatalogStitch-Eval [7] using Occluder PSNR (↑), following their respective evaluation protocols. As shown below, Chameleon achieves the best results across all five metrics.

MethodShadow LR ↓Shadow LS ↑Reflection LR ↓Reflection LS ↑Occ. PSNR ↑
Copy & Paste81.20.09858.70.07816.5
TF-ICON82.50.07358.90.03915.2
TALE79.80.09258.60.04220.0
PrimeComposer76.40.07363.50.03212.6
AIComposer79.90.08758.70.04219.4
Ours71.60.15956.20.09923.3

Qualitative comparisons in Fig. 14 further show more faithful shadows and reflections, as well as better preservation of foreground-background occlusion relationships. We have added these analyses and evaluations to Appendix A.4 (Figs. 13–14 and Tab. 10), with revisions marked in red.

References.

[1] Khosla et al. Supervised Contrastive Learning. NeurIPS 2020.

[2] Robinson et al. Contrastive Learning with Hard Negative Samples. ICLR 2021.

[3] Zhu et al. Less is More: Masking Elements in Image Condition Features Avoids Content Leakages in Style Transfer Diffusion Models. ICLR 2025.

[4] Xu et al. StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer. CVPR 2025.

[5] Liu et al. Shadow Generation for Composite Image Using Diffusion Model. CVPR 2024.

[6] Zhao et al. Reflection Generation for Composite Image Using Diffusion Model. ICME 2026.

[7] Jain et al. CatalogStitch: Dimension-Aware and Occlusion-Preserving Object Compositing for Catalog Image Generation. CVPR 2026 Workshop (HiGen).

Reviewer_rn9KOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses cross-domain object compositing, where a foreground object from one visual domain is inserted into a background from another domain while preserving its identity and adapting its style. To provide training data, the authors build ChameleonDataset through a reverse pipeline that starts from real stylized composite images, segments their objects, removes them to obtain backgrounds, and generates foreground references, yielding about 200K training samples together with an evaluation benchmark. The proposed framework, Chameleon, consists of two stages. ChameleonEncoder is first trained on DINOv3 with Signed Style-Content Graph Learning (SSCGL) to disentangle style and content tokens, and these tokens are then injected into a DiT with Style Gap-Aware Anisotropic Modulation (SGAM), which regulates style attention according to the foreground-background style gap, spatial region, and diffusion timestep. Experiments on the TF-ICON and AIComposer benchmarks, the proposed benchmark, and a user study show improvements over in-domain, cross-domain, cascaded, and commercial baselines.

Critical strengths
  • The paper addresses the problems of cross-domain image composition through a two-stage framework, in which ChameleonEncoder is first trained with SSCGL to disentangle style and content, and its tokens are then injected into a DiT with SGAM. In particular, the proposed ChameleonDataset is likely to be highly valuable for future research on cross-domain composition, where training data are scarce. The reverse pipeline, which starts from real stylized composites and synthesizes only the input conditions, is a good idea for a setting where real paired data are difficult to obtain.

  • The architecture is novel in that it handles failure cases that simple stylization followed by insertion or harmonization cannot resolve, such as occlusion, missing reflections, over-stylization, and ignored mask regions (Fig. 1). SSCGL is also effective for style-content disentanglement. Rather than relying on augmentation-based positive pairs, it assigns opposite signs to the same image pair across the style and content spaces, which yields clearly separated style clusters (Fig. 5).

Critical weaknesses

W1. Accuracy of VLM-based data filtering.

The paper reports the thresholds (a keep score of at least 70 and recoverability of at least 0.35) and the retention rate (82.1%, Appendix A.4), but not how well the filter agrees with human judgment. A human verification on a sampled subset would help.

W2. Reference generation.

The reference generation step lacks detail. The paper only states that Qwen-Image-Edit is used with CLIP-I filtering, without describing the prompt or the target style, and the generated references in Fig. 2 appear to be photorealistic. For cross-domain composition, it seems more natural to generate references in diverse domain styles beyond the source image rather than only real images, especially since the paper claims generalization to stylized-to-stylized compositing (Appendix G.2).

W3. CSD and over-stylization.

In the CSD metric, StyleSSP + InsertAnything achieves the highest scores on both benchmarks (0.5535 and 0.4898 in Table 2), which the paper attributes to over-stylization based on Fig. 1(g). The degree of over-stylization in this cascaded baseline is therefore likely to depend on the performance of StyleSSP, and it should be verified with more diverse style transfer methods whether over-stylization occurs regardless of the stylization method. That said, a cascaded two-stage pipeline is admittedly less practical than the proposed method.

Minor, but should be fixed

In Fig. 6, the encoder is labeled as Sec. 3.2.1, which does not exist. It should be Sec. 3.3.

Questions for authors

Please refer to "Critical weaknesses sections".

8 comments
AuthorsOpenReview ↗

Response to Reviewer rn9K (Part 1/2)

We sincerely thank Reviewer rn9K for the careful review and constructive feedback, and for recognizing the contributions of our dataset construction and framework design. We address each concern below and summarize the corresponding changes in the revised manuscript (marked in red).

[W1] Accuracy of VLM-based data filtering.

We agree that human verification is important for validating our VLM-based filtering. We evaluated 400 candidates (200 VLM Keep and 200 VLM Reject), independently assessed by five annotators using the same filtering criteria (Appendix H.2), without access to VLM predictions.

Human agreement was 90.5% among VLM Keep candidates and 42.5% among VLM Reject candidates. Notably, the disagreement rate was substantially higher for VLM Reject decisions (57.5%) than for VLM Keep decisions (9.5%), suggesting that the VLM applies more stringent selection criteria than human annotators, prioritizing the quality of retained training samples. Weighting the class-conditional agreement rates by the original VLM Keep/Reject distribution (82.1%/17.9%) yields an estimated overall agreement of 81.9%, with 83.2% human inter-annotator agreement. We added the evaluation protocol (Fig. 10), quantitative results (Tab. 8), and representative agreement/disagreement cases (Fig. 11) to Appendix A.4.

[W2] Reference generation.

As the reviewer correctly observed, our original model uses photorealistic foreground references, generated by Qwen-Image-Edit [1] with prompts explicitly requesting realistic photography while preserving object identity. We provide the prompt templates and generation details in Appendix A.4 (L1283–1300).

We agree that incorporating diverse stylized references is a valuable suggestion. We conducted a controlled experiment using a 1:1 mixture of photorealistic and stylized references, randomly sampling from 12 styles independently of the source images. We changed only the style-specific prompt instructions, keeping the dataset size and training configuration unchanged (L1301–1303).

Training referencesReal-to-StyleStyle-to-Style (Seen)Style-to-Style (Unseen)
CLIP-I ↑CSD ↑CLIP-I ↑CSD ↑CLIP-I ↑CSD ↑
Photorealistic only0.8420.4630.8630.4560.8480.451
Mixed (1:1)0.8270.4580.8530.4620.8540.439

The results (Tab. 9) show comparable performance in CLIP-I (identity preservation) and CSD (style consistency) across the three settings, with mixed differences (at most 0.015 in CLIP-I and 0.012 in CSD). Qualitative results (Fig. 12) further illustrate that the original model produces visually coherent Style-to-Style composites with diverse stylized references, despite being trained exclusively with photorealistic references.

We believe this behavior is consistent with our role-specific conditioning: ChameleonEncoder, trained with SSCGL, provides foreground content tokens CT(If)C_T(I_f) to preserve object identity and background style tokens ST(Ib)S_T(I_b) to guide style adaptation. This separation may allow the model to retain identity-relevant information even from stylized foreground references while adapting their appearance to the background style (Secs. 3.3–3.4). We appreciate this valuable suggestion and have added the reference generation prompts, controlled experiment, and qualitative analysis to Appendix A.4 of the revised manuscript (marked in red).

AuthorsOpenReview ↗

Response to Reviewer rn9K (Part 2/2)

[W3] CSD and over-stylization.

We appreciate the reviewer's constructive suggestion to evaluate more diverse style-transfer methods and further substantiate our observations. Accordingly, we extended our evaluation to eight style-transfer methods combined with three insertion models (24 cascaded pipelines) on ChameleonDataset$_{ev}$, including recent approaches such as StyleGallery [2], StyleExpert [3], and USO [4] (CVPR'26).

The extended quantitative results reveal a consistent trade-off: across all 24 combinations, CSD increases while CLIP-I decreases relative to the corresponding insertion-only baselines. Below, we summarize the results using InsertAnything (IA; AAAI'26) [5] as the insertion model. Full results for all 24 combinations across five metrics are provided in Tab. 28 (Appendix F.3).

MethodLPIPS ↓CLIP-I ↑CSD ↑
StyleID [6] + IA0.69880.73390.4305
CSGO [7] + IA0.72750.71230.5022
StyleShot [8] + IA0.73280.70570.5156
StyleSSP [9] + IA0.73780.71490.5322
Attention Distillation [10] + IA0.72330.71030.5250
USO [4] + IA0.71920.73260.4556
StyleGallery [2] + IA0.72240.72800.4631
StyleExpert [3] + IA0.72180.72980.4243
Ours0.46510.84250.4629

As the reviewer correctly notes, StyleSSP + InsertAnything achieves the highest CSD in Tab. 2, despite the over-stylization shown in Fig. 1(g). This illustrates that higher CSD can reflect excessive stylization rather than better compositing. In contrast, Ours achieves consistent stylization while preserving foreground identity (CLIP-I 0.8425). Qualitative comparisons (Fig. 26) show that over-stylization and identity degradation occur across diverse style-transfer methods, rather than being specific to StyleSSP.

To further characterize these limitations, we identify two representative failure modes (Fig. 27). (1) Error propagation from stylization occurs when foreground distortions persist through subsequent insertion, whereas (2) insertion-stage failures arise when models such as Dreamfuse [11] fail to generate the foreground even without prior stylization. Both can compromise the final composite. We have added the extended quantitative comparison (Tab. 28, Appendix F.3) and qualitative analyses (Figs. 26–27, Appendix G.2) to the revised manuscript (marked in red).

[Minor] Fig. 6 label.

Thank you for pointing this out. We have corrected the label from Sec. 3.2.1 to Sec. 3.3 in Fig. 6 of the revised manuscript, with the change marked in red.

References.

[1] Wu et al.: "Qwen-Image Technical Report." arXiv (2025)

[2] He et al.: "StyleGallery: Training-free and Semantic-aware Personalized Style Transfer from Arbitrary Image References." CVPR (2026)

[3] Zhu et al.: "Mixture of Style Experts for Diverse Image Stylization." CVPR (2026)

[4] Wu et al.: "Unified Customized Generation by Disentangled Reward Modeling." CVPR (2026)

[5] Song et al.: "Insert Anything: Image Insertion via In-Context Editing in DiT." AAAI (2026)

[6] Chung et al.: "Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer." CVPR (2024)

[7] Xing et al.: "CSGO: Content-Style Composition in Text-to-Image Generation." NeurIPS (2025)

[8] Gao et al.: "StyleShot: A Snapshot on Any Style." IEEE TPAMI (2026)

[9] Xu et al.: "StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer." CVPR (2025)

[10] Zhou et al.: "Attention Distillation: A Unified Approach to Visual Characteristics Transfer." CVPR (2025)

[11] Huang et al.: "DreamFuse: Adaptive Image Fusion with Diffusion Transformer." ICCV (2025)

Thank you for the detailed response and the additional experiments.

[W1] I appreciate the human verification of the VLM-based filtering. Since the human agreement on VLM Reject decisions is only 42.5%, a substantial portion of the discarded samples may in fact be valid. Although rejected samples account for only 17.9% of the candidates, it would be helpful to know whether these false rejections are concentrated in specific styles or categories, as such concentration could bias the style coverage of the dataset. (This is just a suggestion and not a requirement.)

[W2] Thank you for providing the reference generation details and the controlled experiment with mixed references. This addresses my concern well.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.