- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
CopyCat proposes a lightweight refinement procedure for pretrained subject-to-image models. It adds a rank-4 Fine-grained Consistency LoRA (FCLoRA), freezes the existing model, and performs four optimization steps using one proxy image as both the reference and reconstruction target, with an empty text prompt. The authors report approximately three seconds of refinement, excluding model loading, after which the model is reused across other subjects without further optimization. Separately, the paper studies removing LoRA from dedicated text-stream modules while retaining image-stream and shared-block adaptation. Experiments on DreamBench and XVerseBench evaluate subject consistency across four pretrained models, alongside a separate retraining study of Visual-only LoRA and supplementary proxy-image comparisons.
Critical strengths
-
A practical refinement procedure with very low adaptation cost. Updating a small adapter using one proxy image and four optimization steps is an attractive way to improve existing models. Reusing the refined model across subsequent subjects avoids repeated subject-specific training, giving the approach practical value if its improvements are reliable.
-
Consistent subject-similarity gains in the reported before-and-after comparisons. Tables 1–3 show improvements across four models and both single- and multi-subject settings. For example, UNO’s single-subject DreamBench DINO score increases from 0.754 to 0.839, and UMO’s single-subject XVerseBench identity score increases from 76.22 to 85.01. These results support the underlying empirical observation across several existing models.
Critical weaknesses
-
Main concern: missing comparisons with contemporary general-purpose image models. The evaluation omits systems such as GPT Image-2.5 and Nano Banana2, which already support reference preservation and identity-consistent editing. Matched comparisons using the same references and semantically equivalent instructions are needed to establish what important limitations CopyCat addresses beyond these systems. The authors should evaluate fine-grained fidelity, instruction following, and image quality, and clarify whether the practical advantage lies in quality, cost, or deployment flexibility.
-
Insufficient documentation of model-specific prompting. The figures show short, generic instructions, but the actual prompt templates and reference-image formatting used for each baseline are not clearly documented. This makes the fairness of cross-model comparisons difficult to assess. The authors should disclose the complete inputs, follow each model’s recommended conventions, and test whether the conclusions persist with explicit reference-preserving instructions while keeping the requested content equivalent.
-
The proxy-to-evaluation image protocol needs clarification. Appendix D names the model-specific proxies, and Figure 2 illustrates a reconstruction example. However, the exact source images used across experiments and whether proxy identities are excluded from DreamBench evaluation are not clearly documented. Explicitly identifying the refinement images and their relationship to evaluation subjects would make the claimed transfer to unseen subjects verifiable.
-
Insufficient direct evidence of fine-grained detail preservation. The reported identity and image-similarity metrics establish aggregate similarity gains but do not directly measure the preservation of specific local details. Figure 4’s face crops illustrate facial identity, while Figure 5 illustrates reduced identity confusion; neither systematically identifies which fine details improve. Annotated local comparisons and targeted quantitative evaluation or a structured human study are needed to substantiate the central fine-grained consistency claim.
Questions for authors
My questions are incorporated into the four points in Critical Weaknesses. Please respond to them individually. Additional comparisons with contemporary image models, clarification of the prompting and proxy-to-evaluation protocols, and stronger evidence of fine-grained detail preservation could materially affect my assessment.
2 comments
The authors' response does not resolve my main concern regarding the paper's novelty and significance. In my view, the core contribution essentially amounts to training an additional lightweight LoRA adapter on top of an existing LoRA-based model, without sufficient methodological innovation. While the additional experiments demonstrate empirical improvements, they do not fundamentally address this concern. Therefore, I maintain my original rating.
Thank you for recognizing the efficiency, cross-subject reuse, and subject-consistency improvements of CopyCat.
1. Comparison with contemporary models
General-purpose image systems target broader capabilities and have different training scales and computational budgets. Their end-to-end training cost is not directly comparable to our additional refinement cost. Our focus is improving the fine-grained subject consistency of existing S2I models, evaluated on established S2I benchmarks. CopyCat requires one proxy image and approximately three seconds of refinement training, followed by reuse across subjects.
We have revised the manuscript to include CRAFT (CRAFT: Constrained Reward via Attention Fine-Tuning for Subject Personalization without Composed Targets, SIGGRAPH Asia 2026), a strong baseline with recently released code and weights. UMO+CopyCat achieves higher single-subject ID/IP on XVerseBench (85.01/84.20 vs. 83.28/83.40) and DINO/CLIP-I on DreamBench (0.865/0.893 vs. 0.809/0.861). Tables 1–3 and Figure 3 report the added comparisons.
2. Prompts and reference formatting
All methods use the same generation requests and original reference images. We follow each method’s official preprocessing and inference settings, adapting the text only for its reference-addressing conventions:
ipfor animal/object examples).the animal/person/subject from image N.For example,
'a cat wearing a rainbow scarf'becomes'the animal from image 1 wearing a rainbow scarf'for CRAFT. The requested content is unchanged. Within each baseline/+CopyCat pair, prompts and reference preprocessing are identical. For the specialized S2I backbones, reference preservation is already a training objective implemented through reference conditioning. Their scene-description prompts do not serve the same role as open-ended editing instructions in a general-purpose system such as GPT Image. An extra “preserve reference details” instruction does not provide new reference information or directly enforce finer correspondence. We therefore follow the native reference-conditioning protocol. We will organize all images shown in the paper and their complete model-specific prompts in the supplementary material.3. Proxy overlap
The original DreamBench evaluation includes the proxy subject. We additionally evaluated UMO and its cat2-refined counterpart after excluding cat2.
Excluding cat2 does not change the conclusion: DINO and CLIP-I improve in both settings. The gains are therefore not explained by including the training proxy.
4. Direct evidence of fine-grained preservation
We added Figures 7 and 8 in the revised manuscript. Figure 7 compares UMO and UMO+CopyCat on a non-proxy dog under three prompts and three seeds per prompt. It shows muzzle and mane details alongside requested pose and appearance changes.
Figure 8 reports a pairwise evaluation using GPT-6.1-sol with medium reasoning effort. Method names were hidden and A/B positions randomized. Prompt adherence was judged before revealing the reference. Fine-grained consistency was then assessed using texture, color, material, shape, parts, and accessories, while ignoring background and pose and respecting prompt-requested changes.
CopyCat win / tie / loss rates (%) against each baseline:
All four models receive more fine-grained consistency wins than losses, while prompt-adherence judgments are predominantly ties. This provides targeted evidence beyond aggregate image-similarity scores.