acceptodds
Withdrawn from ICLR 2027

TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When using most generative image editors to edit an image over multiple turns, minor unintended changes may accumulate, resulting in a corrupted result in the end. This long-term editing drift is normally mitigated through heavy large-scale pretraining or additional modules that may introduce inference overhead. In this work, we introduce TailSteer, a simple post-training framework that improves the multi-turn robustness of pretrained editors without changing their architecture or introducing additional computational overhead. TailSteer follows a simple principle: drift is often introduced in the last few denoising steps. Therefore, it keeps earlier denoising dynamics mostly unchanged and corrects only the tail of the diffusion trajectory to preserve fine-grained details in the source image. These corrections are learned from the editor's own denoising rollouts, using unchanged source content as a preservation reference rather than requiring curated ground-truth editing pairs. Experiments on three backbones (FLUX 1, FLUX 2, and SenseNova families) show that TailSteer can consistently reduce cumulative non-target drift: after TailSteer fine-tuning, the models can normally support 5–10 editing turns without significant errors, while the original models often produce obvious artifacts in fewer than five turns.

Primary area:
generative models
Keywords:
image editing, multi-turn image editing, generative models, diffusion models, flow matching, image consistency

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_CCXYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TailSteer, a post-training framework for reducing cumulative drift in multi-turn image editing. By leveraging the pretrained editor's own denoising rollouts, TailSteer progressively steers the trajectory tail toward preservation-aware targets while maintaining the original dynamics at earlier steps.

Critical strengths
  1. TailSteer introduces a straightforward trajectory-level refinement strategy that improves multi-turn preservation without architectural modifications or additional inference overhead.
  2. The framework constructs supervision from the base model's own rollouts, avoiding the need for manually curated ground-truth editing pairs.
Critical weaknesses
  1. This work shows improved non-target preservation across editing turns, it does not quantitatively evaluate whether the editing instructions are successfully executed at each turn. The current evaluation primarily relies on masked PSNR for multi-turn editing and single-turn benchmarks for editing capability. And, the quantitative multi-turn evaluation in Figure 8 is based on only eight 10-turn editing sequences per backbone, it is a limited scale.
  2. Although TailSteer improves multi-turn preservation, Table 1 reveals noticeable degradation in certain single-turn editing capabilities, particularly on RISEBench. And the paper does not quantitatively evaluate multi-turn instruction-following success, making it difficult to determine whether improved preservation is achieved without compromising the execution of editing instructions.
  3. How this work constructs masks from the differences between source images and base-model outputs? Since these differences may contain both intended edits and unintended visual drift, the resulting masks could incorrectly classify unintended changes as editable regions, thereby preserving such errors in the supervision targets.
Questions for authors
  1. Could the authors provide quantitative evaluations of multi-turn instruction-following success to demonstrate that improved preservation does not come at the expense of editing capability?
  2. How are preservation masks constructed from the differences between source images and base-model outputs?
  3. How much do the identity-editing and local-editing cached trajectories individually contribute to the performance gains? Could the authors provide an ablation study comparing training with each data type separately and their combination?
  4. This work's multi-turn examples focus on object addition and localized modifications. Could the authors evaluate TailSteer on more diverse editing operations, such as object removal, replacement, attribute modification, and spatial manipulation, to demonstrate its robustness across different types of multi-turn editing tasks?
Reviewer_VK3QOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TailSteer post-trains image editors to reduce accumulated drift over repeated edits. Preservation-aware targets combine source content in unchanged regions with the base editor's chosen local edits. Training matches the base velocity early and steers toward the target in the trajectory tail, using short student rollouts from cached base states. Inference uses only the refined model with its original sampling schedule. Validation covers three backbones, single-turn GEdit/RISE/ImgEdit benchmarks, qualitative training ablations, and masked source PSNR over eight ten-turn sequences, reporting improved non-target preservation.

Critical strengths
  1. A practical preservation mechanism without extra inference components. The endpoint retains the base editor's chosen local edit while restoring source content in preservation regions. Early teacher matching and tail correction train one refined model, which then runs with the original inference schedule. This is a useful deployment property.
  2. Positive evidence across three backbones. Figure 8 reports reduced non-target drift, and Table 1 evaluates single-turn capability on several benchmarks. Basic training settings and qualitative component comparisons are provided. These results make the proposal more substantive than an untested training idea, although they do not settle the central joint-outcome question.
Critical weaknesses
  1. The central multi-turn claim lacks joint success/preservation evidence. Figure 8 reports masked source PSNR over eight ten-turn sequences (p.9), without quantitative instruction success, retention of earlier edits, or multi-turn competitor results. Changing less can improve this metric without improving editing. Single-turn results do not resolve the ambiguity: SenseNova's RISE spatial accuracy falls from 46 to 37/37/30 and overall accuracy from 31.39 to 29.44/26.67/26.39 across the three settings (Table 1). The claimed capability-retention trade-off therefore remains insufficiently established.
  2. The training contribution cannot yet be evaluated or reproduced adequately. Figure 7's component comparisons are qualitative rather than matched-budget quantitative results. The exact blending schedule is deferred to an appendix (p.7, lines 363–364), but the supplied PDF ends at p.12 and no separate supplement is supplied. General difference-based mask construction is described; precise thresholds/filtering rules and evaluation sequence/split definitions are not. These gaps prevent assessment of whether the specific training design produces the reported advantage over simpler alternatives.
Questions for authors
  1. Assessment-critical analysis of existing outputs: On the evaluated sequences, how often are current instructions fulfilled and previous edits retained alongside improved background PSNR? Please provide sequence-level uncertainty and any available matched multi-turn comparator results.
  2. Experimental clarification: Can Figure 7's training controls be quantified under matched data and training budgets? Was FreqEdit's late-stage quality guidance enabled in the reported comparison, and what explains the RISE declines?
  3. Reproducibility clarification: Can the referenced supplement provide the exact blending schedule, mask/filter thresholds, and evaluation sequences with train/test separation? This request concerns the supplied materials, not an assertion that the appendix does not exist elsewhere.
Reviewer_GvrAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TailSteer, which aim to reduce the error accumulation in multi-turn image editing. It aims to construct training targets that retain the model’s intended edits while preserve the unchanged content. The authors trian the model with model's own rollout states to preserve early denoising trajectory. But change the final steps to correct the shift content. Experiments on three different models show improved content preservation with largely maintained single-turn editing performance.

Critical strengths
  • The paper is well written and easy to follow.
  • The experiments were conducted in serveral base model which suggests the generalization of the proposed method.
  • The results shows the strong preservation of the unchanged region after multi-turn editing. It indicates the effectiveness of the proposed method.
Critical weaknesses
  • One of my concern is the training data quality. The input of the model is from online distribution while the editing target is constructed by a mask-based image-merging. (line 248). The constructed target is not ensured to be smooth at the mask boundaries. If there could be some "visual artifacts" in the training data, the output of the model will also generate images with artifacts.
  • The following question is that: if the source of the improved preservation coms from the "merged image", is it neccessary for us to train a model to learn this merging? We can usethis "merging" operation for each generated image from the original model without training.
Questions for authors

There is no question of this paper.

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

Related papers

Discussion (0)

Sign in to comment.