acceptodds
Withdrawn from ICLR 2027

CSTRA-OPD: Counterfactual Spatio-Temporal Alignment for Region-aware On-Policy Distillation in Local Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large instruction-based image editors can teach compact models, but local editing requires changing the target region while preserving its surroundings. Standard on-policy distillation (OPD) matches teacher and student at student-visited states, yet a global teacher target may propagate unintended changes outside the edit region. Uniform sampling across denoising states also ignores differences in their influence on the final edit. We introduce CSTAR-OPD, a Counterfactual Spatio-Temporal Alignment framework for Region-aware OPD. At the same student-visited state, we evaluate a frozen teacher under both the edit instruction and an identity instruction. The student matches the edit-conditioned velocity inside the target mask and the identity-conditioned velocity outside it, using separate region-normalized losses so that the background does not dominate the objective solely by area. The teacher and mask are used only during training; inference requires neither. Analyses of semantic direction and 4B-to-9B model switching motivate sampling one student state per example from states 0-13 of a 50-step schedule. On GIE-Bench, CSTAR-OPD reduces outside-mask M-MSE by 17.7% relative to full-range global OPD while maintaining the same 88.98% functional correctness. These results support combining early-state supervision with region-specific teacher targets to balance editing and preservation.

Primary area:
generative models
Keywords:
Local Image Editing, On-Policy Distillation, Diffusion Models, Counterfactual Supervision, Region-Aware Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kMudOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a region-aware on-policy distillation method for local image editing. The method uses an edit-conditioned teacher target inside the editing mask and an identity-conditioned preservation target outside the mask, while additionally restricting teacher supervision to an empirically selected early denoising window.

Critical strengths

(1) The paper aim to address a practical problem in local image-editing distillation: transferring editing capability from a stronger teacher while avoiding unintended changes in non-target regions.

(2) The region-aware supervision is simple and practically appealing: different teacher targets are used inside and outside the edit mask during training, while no mask or teacher is required at inference.

Critical weaknesses

(1) My main reservation is that the temporal component of the proposed method is not convincingly supported by the training results. A central part of the paper is the claim that OPD supervision should be concentrated in the early denoising stage, motivated by the trajectory and model-switching analyses. However, the ablation in Table 2 shows that, with the same region-aware supervision and one sampled state, full-range sampling (0–49) actually achieves better preservation than early-only sampling (0–13), while maintaining the same functional correctness. The main demonstrated advantage of the early window is therefore reduced student rollout length rather than improved editing quality. Moreover, the appendix itself notes that the switching analysis does not establish that early-only OPD is optimal and that a matched-budget comparison is needed. Given that “spatio-temporal alignment” is presented as the central contribution, I do not find the current evidence sufficient to support the temporal part of the claim.

(2) I am also not yet convinced that the empirical evaluation is sufficient to establish the significance and generality of this rather specific modification to OPD. The main effective change is to use an edit-conditioned teacher inside the mask and an identity-conditioned teacher outside it, but the evaluation is limited to a single FLUX 9B-to-4B teacher–student pair and is mainly compared with a relatively simple Base OPD baseline. The related-work section discusses several recent OPD approaches, yet none is included in the main comparison. Since the proposed method is fairly specialized, stronger comparisons and evidence that the gain persists beyond this particular model pair would, in my view, be important for demonstrating a sufficiently broad contribution for ICLR.

Questions for authors

(1) Why is the temporal window fixed to states 0–13? Could the authors provide matched-budget comparisons with middle, late, and full-range sampling?

(2) How are the training masks obtained? How robust is the method to imperfect masks?

(3) Can the authors compare with some more recent OPD baselines mentioned in the related works section beyond Base OPD?

(4) How well does the method generalize beyond the FLUX 9B-to-4B setting? Does it require closely matched teacher/student architectures or latent spaces?

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

This paper studies teacher-to-student distillation for local image editing, focusing on unintended changes outside the requested edit region. At the same student-visited denoising state, the teacher is queried with an edit instruction and an identity instruction. The student matches the edit field inside the mask and the identity field outside it, with separately normalized regional losses. Training samples one supervision state from states 0–13 of a 50-step trajectory; inference requires neither the teacher nor a mask. Experiments use FLUX.2 Klein 9B/4B, with training on four CoCoEdit task categories and evaluation on 599 GIE-Bench examples. Relative to full-range global OPD, the method reduces outside-mask M-MSE from 791.31 to 651.40 while retaining 88.98% functional correctness.

Critical strengths

Local editing requires both making the requested change and preserving unrelated content, whereas global teacher matching does not distinguish these objectives. Constructing two conditional targets at the same student state and assigning them to different regions is a well-motivated response. Using the teacher and mask only during training also makes the objective applicable to a student whose inference interface does not take a mask.

Table 2 provides more than a comparison between the final model and Base OPD. With full-range sampling and one state held fixed, adding the regional objective reduces M-MSE from 791.31 to 622.32 and raises M-SSIM from 0.9199 to 0.9267, while functional correctness remains at 88.98%.

Critical weaknesses

Early-window sampling is a main contribution, but the full-range regional variant in Table 2 achieves better preservation on every reported metric at the same 88.98% functional correctness; its M-MSE is 622.32 versus 651.40 for the final configuration. The paper acknowledges this, so my concern is how the proposed trade-off is justified.

Moving from global OPD to the regional objective changes two factors at once: the background target becomes identity-conditioned, and the edit and background losses are normalized separately. The latter already changes the training weight assigned to small edit regions. The Region toggle in Table 2 therefore supports the combined objective, but does not show how much of the gain comes from the counterfactual teacher target.A key control would retain exactly the same regional weighting and normalization while matching the edit-conditioned teacher in both regions. An edit-region-only variant would also help determine whether the identity branch supplies a distinct preservation benefit. This distinction matters to the assessment of the paper’s central methodological contribution.

Questions for authors

1.With the sampling window, regional normalization, and training steps fixed, what happens if both regions use the edit-conditioned teacher?

2.Can the authors report time per update, total GPU-hours, and results under matched total compute for early and full-range regional supervision, including the extra identity-conditioned teacher branch?

3.Are the masks for the 21,000 training examples supplied by the dataset, predicted by a segmentation model, or generated separately? How is the target region defined for object addition?

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

The paper proposes CSTAR-OPD for distilling a 9B image editor into a 4B student. At student-visited states, the same teacher provides edit-conditioned predictions inside the edit mask and identity-conditioned predictions outside it. Training uses separately normalized regional losses and samples supervision from early states.

Critical strengths

The regional objective improves background preservation under both early and full-range sampling without reducing aggregate editing correctness accroding to the ablation. The method uses masks during training while keeping inference mask-free, and can indeed improve the editing quality in some cases.

Critical weaknesses
  • The method is useful, but the contribution appears incremental. Its benefits are demonstrated in a narrow local-editing setting, and the current comparisons do not establish a clear advantage over simpler region-weighted OPD.

  • The method mainly modifies teacher supervision within an existing OPD framework through region-specific targets, loss normalization, and early-state sampling. These changes are useful, but the methodological novelty appears limited. It seems more like a practical adjustment to compensate for an imperfect teacher model than a substantial advance in OPD.

  • How are training masks defined when valid edits extend beyond the object, such as changes to shadows or reflections? Although the loss supports soft weights, it is unclear how these masks are obtained and whether preservation supervision incorrectly penalizes necessary changes outside the annotated region. If we see in the first example of Figure 5, the added light appears locally inserted, with little visible effect on its surroundings. Lighting changes may legitimately extend beyond the edit mask. How does the method accommodate such changes, and could the outside-mask metrics penalize physically plausible results?

Questions for authors
  • The red overlays in manu figures appear misaligned with the target objects. Figure 5, particularly for the lamp and smartphone examples. Do these represent the evaluation masks, or is this a visualization issue?

Minor Issue: In figure 4, "Forzen" should be "Frozen"; "euqation" should be "equation"

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

Related papers

Discussion (0)

Sign in to comment.