acceptodds
Withdrawn from ICLR 2027

CRAudit: Predicting, Preventing, and Repairing Conditional Response Collapse in One-Step Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

One-step distilled text-to-image models are fast, but often collapse distinct editing instructions into nearly identical responses. The central challenge is to identify vulnerable text responses early to direct resources toward their preservation or repair. We address this gap through the conditional response operator, the Jacobian from text-embedding perturbations to predicted image-residual changes. Conditional Controllability Spectra (CCS) audits this operator in a fixed text-derived basis and exposes contraction of weak response modes. Across four distillation families, CCS effective rank tracks an editing-quality drop. Using only the first three of 15 checkpoints, terminal-risk prediction yields Brier 0.088 and AUROC 0.89; threshold-rank prediction correlates at . The same aligned coordinates guide spectrum-preserving distillation (SPD) toward at-risk modes and spectral residual adaptation (SRA) toward frozen-student repair with 0.54% additional parameters. On PIE-bench, SPD+SRA reaches 25.06 dB PSNR and 24.58 CLIP-E in 0.35 s; it obtains 75.3% participant-clustered preference over FlowEdit and 71.7% in an opponent-balanced tournament (). Compression audits characterize contraction; interventions test selected response directions. Matched allocation yields 0.49/0.47 CLIP-E gains for SPD/SRA; fixed-architecture controls further test predictive selection, separating allocation from adapter capacity.

Primary area:
generative models
Keywords:
One-Step Image Editing, text-to-image generator, Fast generation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_BxfQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces CRAudit, a method to find and fix "conditional response collapse" in one-step image editors. After distillation, different editing instruction sgive almost the same result. Instead of only checking final output's fidelity or so, CRAudit pushes the text embedding along a fixed set of semantic directions and measures how much the edit made by those distilled model changes.

The authors argue that this collpse should be predicted early so they forecast from the first three model checkpoints which directions will fail. During distillation they add a loss which asks the one-step model to change the image as much as the original model does when the text embedding is moved along each direction, and this loss puts more weight on the directions that are forecast to fail. For a model that is already distilled they keep it frozen and train a small adapter that brings the weak directions back. The authors show that edit quality drops fast once only a few directions still change the output. Their final editor is fast and people prefer it over FlowEdit, which is a strong editing baseline.

In conclusion, this paper mainly shows:

  • The loss of editing ability in one-step models shows up in the spectrum of the text-to-image response.
  • The weak tail directions carry edit control for specific categories because changing them moves only the target edit type.
  • Protecting and repairing these directions gives a fast editor with strong quality and human preference.
Critical strengths
  • The paper targets an important behavior of distilled editors. One-step models often ignore an edit and return the source image. The paper turns this into a measurable collapse of the model's response to text. This adds a mechanism to the diversity loss [1] and the weak edits [2] already reported for fast models.
  • The effective-rank finding gives the field a method to measure this collapse as the rank of the text response falls during distillation. Especially they show that edit quality drops sharply once the rank passes a critical value in all four model families. The weak tail directions also carry edit control for specific categories.
  • Each proposed component have comprehensive experiments on them to understand the property of each method or components, although these same experiments withdraw its claims (Weakness 1).

References

  • [1] R. Gandikota and D. Bau. Distilling Diversity and Control in Diffusion Models. WACV 2026
  • [2] G. Deutch et al. TurboEdit: Text-Based Image Editing Using Few-Step Diffusion Models. SIGGRAPH Asia 2024
Critical weaknesses
  1. In the main text, the paper itself dismiss every single main claim by itself after named them. For example: It calls the bound "a post-hoc audit, not a forecast". The predictor's Brier score is worse than a constant baseline, and only two of 32 cases fail. Prediction does not beat current-state allocation after correction. The loss terms are "not established" as necessary. The repair "cannot certify a successful finite edit". Almost every other result is also followed by what it does not establish. So the reader cannot tell what the paper finally claims, and the title's "predicting, preventing, and repairing" is NOT supported as stated.
  2. The claim or components do not connect to each other in method, results or in conclusions. The audit, risk predictor, distillation loss, adapter and a separate repair baseline each have their own terms and their own eval set. The paper keeps their results separated and never shows how they interact to each others. The related work session have similar issue as the paper lists nearby methods without saying what they missed. It also omits the closest prior observations. Distilled models keep control directions but do not activate them [1], and few-step editing gives too weak edits [2]. So the reader cannot see how the rank-collapse finding leads to the editor's gains, which is the link the paper is built on.
  3. Fig. 1 is hard to read and looks unfinished, the lattering is stretched and the plots and labels show visible artifacts. Some labels are pasted over the original text with some stray marks around them. We highly suggest that the authors redo this figure with proper plotting and drawing tools.
  4. The main result are thin and their provenance is still unclear. The protocol states five seeds and bootstrap intervals, but the main table has no intervals reported. SSIM and DreamSim are listed as metrics but never reported. Adding the adapter doesn't change the headline edit score. The repro statement lists figures, deps, sanity checks and measurement record. Secion S10 also describes a release with entry points. But the paper submissions has no supplementary material and I also didn't find any url point the anonymous code repo.

References

  • [1] R. Gandikota and D. Bau. Distilling Diversity and Control in Diffusion Models. WACV 2026
  • [2] G. Deutch et al. TurboEdit: Text-Based Image Editing Using Few-Step Diffusion Models. SIGGRAPH Asia 2024
Questions for authors
  1. Why do the current-energy and current-error arms in Table 4 give exactly the same numbers in every cell? Why does the Brier section use a prevalence of 0.0654 when 2 of 32 series fail?
  2. Which teacher and student does each family use? How is CLIP-E computed? Were all reported numbers produced by new training runs for this submission?
  3. Suppose the paper were centered on the rank-collapse finding. Does the critical rank depend on the edit category or the distillation family? Does the effective rank at an early checkpoint alone allocate the budget as well as the risk predictor?

Besides the questions above, please also refer to the weakness part.

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

This paper addresses conditional response collapse in one-step image editing, where distilled models lose sensitivity to certain text instructions. CRAudit proposes Conditional Controllability Spectra (CCS) to measure changes in text-conditioned image responses and predict which response directions are likely to fail during distillation.

The framework combines three components: CCS for identifying vulnerable directions, Spectrum-Preserving Distillation (SPD) for preventing their degradation, and Spectral Residual Adaptation (SRA) for repairing frozen models. Experiments across four distillation families and PIE-bench demonstrate improvements in editing quality, computational efficiency, and human preference over several existing methods.

Critical strengths

Unified approach to diagnosis, prevention, and repair. The paper connects spectral analysis with practical interventions, using the same response directions to identify, preserve, and repair conditional control. This is an interesting contribution beyond simply measuring editing quality or improving distillation.

The authors evaluate multiple distillation families, compare against established editing and derivative-matching methods, and include ablations, controlled interventions, and human preference studies. The reported editing performance and low inference latency are promising. The matched-allocation experiments also help separate the benefits of selecting response directions from simply increasing model capacity.

Critical weaknesses

Limited methodological novelty and insufficient validation of early prediction. Jacobian matching and spectral preservation have already been explored by Czarnecki et al. (2017), Srinivas and Fleuret (2018), Zhang and Pilanci (2024), and Huang et al. (2026). Although CRAudit introduces prediction-guided interventions, its main predictive evaluation covers only 32 mode series with two failures. This limited sample makes the reported AUROC of 0.89 difficult to assess reliably and weakens the evidence for early prediction over simpler alternatives.

Limited statistical evidence and generalization. The prediction-guided allocation gains are promising but do not remain significant after multiple-comparison correction (Holm-adjusted p = 0.115). Fixed-architecture comparisons also lack confidence intervals. Furthermore, the framework relies on a fixed semantic basis, with limited validation across unseen instructions and architectures. These concerns weaken confidence in its robustness and practical applicability.

Questions for authors

Can the authors validate early failure prediction on a larger independent dataset with more failure cases and comparisons against simpler predictors?

Are the improvements over Jacobian matching, GAD, and standard spectral regularization statistically significant under matched computational budgets?

How well does CRAudit generalize to unseen editing instructions, different semantic bases, and model architectures? What is the additional cost of spectral auditing?

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

This paper studies conditional response collapse in one-step distilled image editing models, where different editing instructions produce similar responses. The proposed CRAudit framework analyzes the conditional response Jacobian within a fixed text semantic subspace using Conditional Controllability Spectra (CCS) and predicts mode failure risks from early checkpoints. These predictions guide Spectrum-Preserving Distillation (SPD) to preserve vulnerable response directions and Spectral Residual Adaptation (SRA) to repair frozen models with a lightweight low-rank adapter. Its main contribution is a shared response coordinate system connecting diagnosis, prediction, prevention, and repair. Validation includes trajectory analysis across four distillation families, early risk prediction, PIE-bench evaluation, ablations, budget- and capacity-matched comparisons, directional interventions, and user studies. Using the first 3 of 15 checkpoints, terminal mode-failure prediction achieves an AUROC of 0.89. SPD+SRA achieves 25.06 dB PSNR and 24.58 CLIP-E, with 0.35-second inference under the reported A100 setup. SRA adds approximately 0.54% parameters. A study with 48 participants reports a 75.3% pairwise preference rate over FlowEdit.

Critical strengths
  1. Diagnostic results can directly guide intervention. Built upon a unified response coordinate system, CRAudit integrates degradation analysis, risk prediction, distillation optimization, and model repair, enabling spectral analysis to deliver practical application value.
  2. The method is validated from comprehensive experimental perspectives. This paper conducts cross-distillation analysis, ablation and matching comparison, directional intervention, and user studies, providing solid evidence from both mechanism interpretation and editing performance to strengthen the persuasiveness of the proposed method.
Critical weaknesses
  1. Limited validation of early failure prediction. The prediction evaluation includes only 32 mode-family series, with two terminal failures, making it difficult to assess the stability of the reported AUROC. Moreover, the Brier score of 0.088 does not establish an advantage over simple constant baselines. Since early prediction connects diagnosis to intervention, uncertainty about its reliability directly affects the strength of the paper’s central contribution.
  2. Insufficient uncertainty analysis of the additional gains from risk-guided selection. Under a fixed architecture, risk-guided selection improves CLIP-E by 0.15–0.22 over alternative strategies, but paired confidence intervals are not reported. A separate family-level comparison does not reach statistical significance after multiple-comparison correction. These results leave uncertainty about whether early prediction provides consistent additional benefits, weakening the evidence for prediction-guided allocation.
Questions for authors
  1. Clarify the reliability of prediction. How are the two endpoint failure samples distributed across the four held-out model families? Could the authors provide uncertainty estimation for the reported AUROC and compare the Brier score with constant prediction baselines under the same evaluation protocol? Such supplements help verify whether early prediction can still deliver reliable information under limited failure samples.
  2. Clarify the benefits brought by predictive selection. For fixed-architecture comparisons, could the authors provide seed-level paired differences and corresponding confidence intervals of risk-guided selection against energy-based, error-based, and random selection baselines? In addition, it is necessary to clarify the relationship between the family-level effects in Table S6 and the aggregated results in Table 4. These analyses help validate whether prediction-guided budget allocation yields stable marginal performance gains.
Reviewer_KZ85OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies conditional response collapse in one-step distilled image editors, where different instructions produce nearly identical edits, and proposes CRAudit to predict, prevent and repair it. It measures how the model’s output responds to a fixed set of semantic directions in text-embedding space and tracks the resulting spectrum (CCS) during distillation. The paper reports that edit quality drops sharply once the spectrum’s effective rank falls below a threshold, and that a Bayesian model fit on early checkpoints can predict which spectral modes will collapse. These predictions guide two methods: SPD, a regularizer that keeps the student’s response close to the teacher’s during distillation, and SRA, a small low-rank adapter that repairs an already distilled model. On PIE-bench, the combined method is reported to beat all compared editors on background preservation, edit alignment and human preference, while also running faster than each of them. However,I am very concerned about the authenticity of the reported results and the experiments. Currently I suggest clear rejection.

Critical strengths

A sensible, low-cost diagnostic tied to a real failure mode.

  • The paper probes how a distilled generator responds to text-embedding directions (64 JVPs in a fixed semantic basis) and tracks aligned modes across checkpoints. This is a cheap and interpretable way to examine a known weakness of one-step editors.
  • Reusing the same coordinates for training-time regularization (SPD) and post-hoc repair (SRA) gives a coherent design.
Critical weaknesses

W1. The main claim, that early prediction helps decide where to spend a fixed intervention budget, is not supported. Without it, the method is close to existing Jacobian-matching and adapter methods.

  • The predictor is no better than a constant guess. Its Brier score (0.088) is worse than always predicting the average failure rate (0.059, a value the paper itself reports). The AUROC of 0.89 is based on only two failures.
  • The allocation gains are not significant. The reported gains (0.49/0.47 CLIP-E) have no seed-level uncertainty, and the paper's own test gives p = 0.115 after Holm correction. In addition, SRA needs early checkpoints from the model's own distillation, which released models usually do not provide.

W2. Several results are inconsistent, so I cannot tell which numbers come from real experiments. The paper itself says that the table values were taken from an author-supplied record and were not independently reproduced (p. 10).

  • Fig. 3 does not look like real data. Its lowest CLIP-E values are about 1–2, while the lowest CLIP-E in any table is about 20.6. Its x-axis ticks are also unevenly spaced, which plotting software would not produce on a linear axis.
  • Other numbers do not add up. The confidence interval for the 75.3% human preference (55 of 73 trials) is narrower than the basic binomial interval for 73 trials, although grouping by participant normally makes it wider. Tables 4 and S7 are said to use a separate dataset that is never described, yet their results for the main SPD model match its PIE-bench results in Table 1 to within 0.01, so it is unclear what data these tables use.
Questions for authors
  1. Please provide per-seed outputs for Tables 1, 3, 4, S3 and S7. What is the intervention set used in Tables 4 and S7 (source, size, overlap with PIE-bench), and which data does each row use? Do the inverse-energy and error orderings of the matched-allocation protocol (Supp. S6.5) coincide? If so, the two Table 4 rows are a single control; if not, why are all four editing metrics identical? Likewise, why do the Table 3 rows without the probe term and without the Gram term report the same effective rank and CLIP-E?
  2. Does it show measured checkpoints or a schematic? How can CLIP-E reach 1–2 when your base and frozen models score 20.6–20.8? Please give per-checkpoint (R, CLIP-E)(R,\ \text{CLIP-E}) values with seeds, and state whether the inset is measured. How was the figure produced, given that its x-axis ticks are unevenly spaced?
  3. Models, editing pipeline and baselines.
    • What is the baseline editor EsE_s that produces xbasex_{\text{base}} (inversion, noise injection, masking)?
    • Which backbone does each Table 1 baseline use?
    • Please add same-backbone rows to Table 1 (the unmodified student with EsE_s, and with Jacobian matching), with seed-level CIs.
    • S10.1 lists SD-Turbo, SDXL-Turbo, PixArt-Σ and SDXL-Lightning as the four architectures, with SD-Turbo as the primary student, whereas Sec. 4.1 and Fig. 3 use SD-Turbo, LCM, DMD2 and SiD. Which models produced Tables 1–4? S9.4 describes cross-architecture transfer but reports only the four-family r=0.94r=0.94; where are the transfer results?
  4. SRA.
    • Is UϕU_\phi an input-independent 16384×816384\times8 matrix? If so, how does a correction confined to a fixed subspace of at most 8 dimensions produce the gains in Table S3?
    • How many modes pass pi>0.5p_i>0.5 per family? With only 2/32 failures, are most of them false positives?
    • How would pip_i be obtained for a released model that has no distillation trajectory?

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

Related papers

Discussion (0)

Sign in to comment.