- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes SelfFix, a self-correction framework that enables a text-to-image generator to correct its own outputs without external LLM/VLM feedback at inference time. Built on SANA-1.0-1.6B, SelfFix freezes the original generator and trains a lightweight adapter with a Content-Aware Correction (CAC) module to determine which image tokens should be preserved or revised. The training data consists of negative-positive image pairs constructed using detector feedback, GPT-4o, and FLUX.1-Kontext-dev. Experiments show that SelfFix improves GenEval from 0.66 to 0.77 with a single correction and achieves 0.91 under a Best-of-20 setting. Additional experiments on Emu Edit and PIE-Bench demonstrate improved image preservation compared to re-generation.
Critical strengths
Interesting and practical problem. The paper addresses an important problem in text-to-image generation: correcting prompt-image inconsistencies without requiring external LLM/VLM feedback or explicit editing instructions at inference time.
Lightweight and effective design. The proposed CAC module provides a simple mechanism for selectively preserving or revising image content. SelfFix trains only 37.5M additional parameters while keeping the original generator unchanged.
Critical weaknesses
-
Limited generalization evaluation. The training data is constructed from approximately 3.9K GenEval-style prompts, which closely match the evaluation categories. Although test prompts are excluded, the strong GenEval results may partly reflect this task-specific training distribution. While Emu Edit and PIE-Bench provide additional evidence, the comparisons are mainly against re-generation. Experiments on more diverse prompts, benchmarks, or base generators would better demonstrate generalization.
-
Unclear fairness and computational cost of the comparisons. Table 2 compares SelfFix against standalone generators with different training data and inference budgets. In particular, the Best-of-20 result (0.91) depends on an additional candidate selector, but the paper does not clearly specify its implementation or computational cost. Although the authors use the same selector across Best-of-N methods, generator calls alone do not provide a complete measure of inference cost. More detailed reporting of the selection procedure and inference costs would make these comparisons more convincing.
-
Insufficient analysis of content preservation in the training pairs. A corrected image should both satisfy the prompt and preserve the parts of the negative image that are already correct; these are distinct requirements. FLUX.1-Kontext-dev may correct an erroneous object while also changing unrelated regions. If such changes appear in the training targets, SelfFix may learn unnecessary modifications. Section 3.1 and Appendix A describe filtering successful corrections using image-quality metrics, but do not report a dedicated quantitative analysis of content preservation in the correction pairs. Such an analysis would help assess the quality of the supervision for conservative correction.
-
The LoRA-only baseline on a frozen base diverges to NaN during training. However, divergence to NaN demonstrates optimization instability and does not by itself establish insufficient model capacity or the necessity of CAC. Because this baseline is important for isolating the contribution of Content-Aware Correction (CAC), the evaluation would be more convincing if the authors demonstrated that this divergence is not an artifact of the chosen training configuration. It would strengthen the empirical conclusions if the authors tested whether this baseline still diverges after adjusting the learning rate, optimizer settings, gradient clipping, initialization scale, or rank.
-
The "Generative Models" and "Image Modification" subsections in Related Work could be combined and shortened. The saved space could be used to present additional experiments or more detailed ablation studies. I would say the paper is well written but lacks sufficient experiments to substantiate its main claims. (Just a suggestion; no response is necessary.)
Questions for authors
Please see the weaknesses above.
4 comments
Response to Reviewer BvC1 (Part 2)
3. Preservation inside the training pairs (W3)
We measured how much each training target changes its negative image. As a reference, we pair each negative with another independent sample of the same prompt.
Table 1. Change from the negative image to its target (medians over pairs, at 512 px).
| Pair | LPIPS ↓ | DINO ↑ | Changed area ↓ |
|---|---|---|---|
| Correction pairs (14,800 training targets) | 0.37 | 0.93 | 34% |
| Re-generation (same prompt, new sample) | 0.67 | 0.75 | 82% |
Targets keep most of the negative image. The training targets keep the content and layout of their negatives (high DINO similarity), far more than an independent sample does. They are not pixel-identical, though, which is why preservation does not rest on the targets alone.
Identity pairs. We add 4,440 identity pairs, in which a correct input must come back unchanged. As Section 4.3 of the paper explains, these pairs teach the model when not to intervene. Within an image that does need revision, CAC keeps the correct regions.
4. The LoRA-only baseline and the role of CAC (W4)
We agree that divergence to NaN does not establish insufficient capacity or the necessity of CAC.
Training stability. The NaN in the paper is what we observed. With the same backbone and training data, (a) diverged in our attempts while SelfFix trained stably. Training without CAC is therefore less robust. To test capacity and the role of CAC directly, we retrained (a) from scratch several times. Under this recipe, (a) completes all 60k steps.
CAC protects correct content. On the same GenEval images, (a) fixes about as many wrong images as SelfFix (32% vs. 34%). But it breaks more correct ones (9% vs. 7%), and its GenEval score is 0.019 lower. Switching CAC off in the trained SelfFix doubles the broken correct images.
5. Related Work (W5)
Thank you for the suggestion. We will merge the "Generative Models" and "Image Modification" subsections and use the space for the analyses above.
Thank you for your response and the additional experiments. They have addressed most of my concerns. However, I will maintain my original score, as I still believe the paper does not meet the acceptance bar for this venue.
Thank you for your kind reply. We are glad that the additional experiments addressed most of your concerns.
Your review raised five specific points, namely generalization beyond the GenEval templates, the fairness and cost of Best-of-20, preservation inside the training pairs, the LoRA-only baseline, and Related Work. Our response answered each of them with new measurements or a concrete revision, and you now consider most of them resolved.
We would therefore be grateful to know which concern remains. Keeping the score without a stated reason leaves us nothing concrete to address. If a specific issue still stands, we would like to address it. If no major issue remains, please reconsider the score in light of the resolved concerns.
Response to Reviewer BvC1 (Part 1)
Dear Reviewer BvC1,
We sincerely thank you for your careful review and for recognizing the practical problem and the lightweight design. Your concerns cover generalization, the cost and fairness of the Best-of-20 comparison, preservation inside the training pairs, and whether the LoRA-only result isolates CAC. We now add a benchmark outside GenEval and specify the selector and its cost. We also measure how much each training target changes its negative image. Finally, we retrain the LoRA-only variant (a) with the same recipe as SelfFix. Below, we address each of your concerns in turn.
We agree that the training prompts follow GenEval templates. They come from Reflect-DiT, the closest prior work, which also improves SANA-1.0-1.6B at test time by conditioning on its previous outputs. We build on its released data, about 3.9k GenEval-style prompts with generated images and detector feedback (Section 3.1 of the paper). SelfFix, Reflect-DiT and SFT are thus trained on the same data and compared on equal terms (the † rows of Table 2). Every test prompt is filtered out, but the categories are the same, which is why we also test beyond the templates.
Evidence outside the templates. Emu Edit and PIE-Bench use captions written for editing, not GenEval templates, and their source images were not generated by SANA. SelfFix receives only the source image and the target caption there. It moves the image toward the caption while keeping far more of the source than re-generation (Table 4 of the paper). The real photographs in Section 4 show the same behavior.
More benchmarks outside GenEval. On T2I-CompBench (2D spatial, 300 prompts, two seeds, official scorer), one correction raises the score from 0.30 to 0.36. Re-generation at the same budget stays at 0.30. The gain is significant in a paired bootstrap over prompts. Other base generators require new failure data and retraining, which we list as the main next step.
Selector. Best-of-N with a verifier is the common form of test-time scaling for text-to-image generators. We follow the protocol of SANA-1.5. All Best-of-N rows that we measured use the same selector, NVILA-Lite-2B-Verifier. It is the public verifier that the SANA authors released for inference-time scaling, and we use it as released. It sees only the prompt and the image and answers "yes" or "no". Candidates are ranked by this answer and its logit, with one pass per candidate. It uses no GenEval feedback, and the GenEval detector scores only the selected image. Reflect-DiT (without †) and SANA-1.5 Best-of-2048 are taken from those papers, as the caption of Table 2 states.
Controlled comparison. The generators at the top of Table 2 of the paper (SDXL, DALLE 3, …) are listed for reference only. The controlled comparison is the SANA-1.0-1.6B block. Every row there uses the same generator, and the † rows also share our training data. At the same budget of 20 generator calls, SANA-1.0-1.6B samples 20 images, while SelfFix samples 10 and corrects each of them once. In both cases, the verifier picks one of the 20 resulting images. With the same verifier, prompts and seeds, the base generator reaches 0.85 and SelfFix 0.91. The selector and the number of calls are identical, so the gain comes from replacing ten new samples with ten self-corrections. The generator judges what in its own image violates the prompt and what to keep. These judgments yield better candidates than sampling again.
Cost beyond generator calls. On one B200 GPU at 1024×1024, a generation takes 0.56 s and a correction 1.52 s. The correction is slower because it adds the current image to the token sequence. The common alternative uses a VLM to diagnose the image and a separate editor to carry out the edit. We measured Qwen2.5-VL-7B as the VLM and FLUX.1 Kontext [dev] as the editor on the same GPU. The diagnosis takes 0.40 s and each edit 13.1 s. SelfFix diagnoses and edits in one pass, so its correction is about nine times faster. These two models add 25.16B parameters on top of the generator. Their weights alone take about 50 GB in 16-bit precision, four times the 12 GB peak memory of SelfFix with the generator included.