acceptodds
Withdrawn from ICLR 2027

Flow-LQR: Flow-Localized Linear Quadratic Regulator for Multi-Turn Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Instruction-based image editors enable iterative visual refinement, but successive edits can accumulate errors and degrade image quality. Our observations indicate that structured degradation arises even when an editor is asked to reconstruct an image without semantic changes, affecting both edited and unedited regions during subsequent editing. Preserving unedited pixels can limit this accumulation, but requires accurate localization, correction within edited regions, and coherent integration of generated and preserved content. We introduce Flow-LQR, a flow-localized linear quadratic regulator for stable multi-turn image editing. Complementary flow responses, calibrated by a reconstruction observation, identify the spatial region of the requested edit. We then reuse the flow predicted by the frozen diffusion transformer as an implicit representation of degradation. Latent residuals at the final denoising step provide disturbance observations in unchanged regions, and harmonic completion extends them into the edited region. A gain-preserving LQR coordinates generated and preserved content while compensating for this disturbance, followed by pixel restoration outside the localized region. We also introduce MTE-Bench for evaluating cumulative degradation over successive edits. Experiments demonstrate superior image preservation over open-source editors and the closed-source editor on both MTE-Bench and Banana100 benchmarks, with improved perceptual quality over the native backbone. Flow-LQR requires no backbone training or external segmentation models, and its disturbance observer reuses existing flow predictions without additional inference costs.

Primary area:
generative models
Keywords:
Multi-turn image editing, Instruction-based image editing, Diffusion transformers, Flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gaZ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Flow-LQR, a training-free and plug-and-play framework for mitigating accumulated degradation in multi-turn image editing. The method first estimates editing mask from flow responses. It then observes latent residuals outside the editing region and uses harmonic completion to infer degradation inside the region. Finally, a gain-preserving LQR coordinates the native editing result with the estimated disturbance, followed by pixel-level restoration of the unchanged regions. The paper also introduces a new ten-turn benchmark MTE-Bench. Experimental results show promising preservation and editing quality.

Critical strengths
  1. The proposed framework is training-free and can be applied to existing flow-based image editing models. The pixel-level restoration also provides a direct way to prevent changes in regions unrelated to the editing instruction.
  2. The paper introduces a new ten-turn editing benchmark and evaluates the method on both MTE-Bench and Banana100. The proposed method achieves strong source fidelity and competitive editing quality.
Critical weaknesses
  1. The technical novelty seems limited. Flow-LQR is mainly a post-processing procedure applied after each editing turn, combining mask estimation, pixel restoration and LQR. In particular, MTC has already introduced LQR control for reducing accumulated errors in multi-turn flow-based editing. The paper does not sufficiently clarify the new technical insight beyond changing the LQR targets and restricting the control locally. Discussion and comparison with model- or trajectory-level multi-turn optimization methods, including training- and RL-based methods, are also insufficient.

  2. The reconstruction motivation is not well connected to the proposed method. The introduction studies degradation under an explicit keep-image reconstruction request. However, this keep-image residual is only used to calibrate the mask. The disturbance supplied to the LQR is instead obtained from residuals in nominally unedited regions during semantic editing. These two residuals are produced under different conditioning and are not necessarily equivalent. The latter may also contain missed semantic changes or global adjustments. The paper provides no analysis showing that the reconstruction degradation observed in the introduction corresponds to the disturbance estimated by the proposed method.

  3. The decomposition of semantic editing and degradation is not sufficiently justified. The method assumes that residuals outside the estimated mask mainly represent degradation and that this degradation can be smoothly extended into the editing region through harmonic completion. However, global lighting, shadows, reflections, or context changes may be valid consequences of an edit. Mask errors can also contaminate the exterior observations. It remains unclear how the method ensures that the completed disturbance does not remove or weaken intended edits.

  4. The method description is difficult to follow and several design choices lack motivation. For example, the role of the analytic reference velocity vrefv_{\mathrm{ref}} is unclear beyond rewriting the endpoint discrepancy. The mask is first estimated in latent space, upsampled for pixel-level compositing, and then downsampled again for latent regulation. Moreover, the method performs pixel compositing, VAE encoding, noise addition, latent LQR refinement, decoding, and another pixel-level restoration. The paper does not explain why this repeated pixel-latent conversion is necessary instead of directly composing and regulating the latent features. It may introduce additional VAE reconstruction errors and boundary artifacts.

  5. The experimental comparison and evaluation protocol are not convincing. Since both Flow-LQR and FreqEdit are plug-and-play methods, the main results should compare native, +FreqEdit, and +Flow-LQR variants under the same Qwen and FLUX.1 Kontext backbones. In addition, the current metrics may reward under-editing, and CLIP-T only evaluates the current instruction. History-aware MLLM evaluation (e.g., metrics in FreqEdit) is needed to assess instruction fulfillment, retention of previous edits, unchanged-object consistency, and boundary artifacts. Besides, reporting only Turns 5 and 10 also provides a limited view of the degradation process, more detailed turn-wise results or complete ten-turn curves would be helpful.

Questions for authors
  1. How sensitive is the method to mask-processing parameters, the candidate gain γ\gamma, the disturbance strength β\beta, and the selected early/late response timesteps?
  2. How does Flow-LQR perform on global editing such as style transfer, where most or all of the image is expected to change and little reliable exterior region remains for disturbance observation? Does the method still provide improvements in this setting?
Reviewer_Vp22OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cumulative degradation in multi-turn instruction-based image editing, where errors introduced at each turn propagate to subsequent turns. The proposed method localizes the requested edit from calibrated flow responses, estimates the disturbance from the final-step latent residual in unedited regions and extends it into the edit by harmonic completion. Also, it integrates the native candidate with the preserved surroundings through a gain-preserving linear quadratic regulator (LQR) before pixel restoration outside the localized region. The paper also introduces a multi-turn image editing benchmark, and reported experiments show improved image preservation over the compared methods.

Critical strengths
  1. Cumulative degradation in multi-turn image editing is a relevant and under evaluated problem, since edited outputs are increasingly reused as inputs. This relevance is the main consideration in favor of acceptance in my assessment.
Critical weaknesses
  1. In Table 3, most of the improvement in LPIPS and SSIM comes from flow-based localization, and adding latent disturbance modeling and gain-preserving regulation changes these metrics only marginally at the final reported turn.

2.The fidelity metrics compare each output with the source image, so they favor methods that edit less, and no measure checks whether each instruction was executed and earlier edits were retained.

  1. Harmonic completion can only extend a smooth disturbance field into the edited region. The compensation can therefore correct low-frequency color drift, but not the over-smoothed or over-sharpened textures that the introduction lists among the degradations. The paper does not test whether the exterior residual predicts the disturbance inside the edit.
Questions for authors
  1. Could the authors report the full method with disturbance compensation disabled under the settings of Table 3?

  2. Could the authors report per-turn instruction execution and retention of earlier edits, and the fidelity metrics restricted to regions that no instruction targets?

Reviewer_baQKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cumulative degradation in multi-turn instruction-based image editing. The key observation is that even a nominal reconstruction, where no semantic change is requested, can introduce structured appearance drift that accumulates over successive editing rounds. To mitigate this problem, the authors propose Flow-LQR, a training-free framework consisting of three components. First, it estimates a soft edit mask from calibrated early- and late-stage flow responses. Second, it treats the final-step latent residual in the presumed unedited region as an observation of generation-induced disturbance and extends this residual into the edited region through harmonic completion. Third, it applies a gain-preserving quadratic regulator that balances the native editing candidate against disturbance compensation, followed by restoration of the input pixels outside the estimated edit support. The paper also introduces MTE-Bench for ten-turn editing evaluation and reports improved source preservation over several open- and closed-source editors.

Critical strengths
  1. The paper targets a practically important and underexplored problem. Error accumulation is a major limitation when an image editor is used iteratively rather than for isolated one-step edits.

  2. The overall design is logically organized. Edit localization, disturbance estimation, and local regulation correspond to three distinct failure modes in multi-turn editing, making the method relatively easy to interpret.

  3. The method is training-free and reuses predictions from the frozen editing backbone for both localization and disturbance observation. It therefore avoids training an additional segmentation or restoration network.

  4. The gain-preserving formulation explicitly prevents an increase in disturbance-removal weight from automatically reducing the coefficient assigned to the native editing candidate. This is a useful design principle for balancing correction and editability.

  5. The paper evaluates recursive editing for up to ten turns and includes component ablations, reference-based fidelity metrics, text-alignment metrics, perceptual-quality metrics, and a human study. The results consistently show substantially stronger preservation of the input image.

Critical weaknesses

1. Generality across editing tasks is not established.

  • The method is most naturally suited to spatially localized edits for which a meaningful unedited exterior exists. It is unclear whether the same assumptions remain valid for global stylization, illumination changes, large background replacement, substantial geometric transformations, or edits that move an object from one location to another. In these settings, a large fraction of the image may legitimately change, and the distinction between edited and unedited regions becomes ambiguous. The paper reports averages over multiple instruction types, but does not provide a category-wise analysis that would reveal where Flow-LQR helps or fails.
  • TODO: The claimed generality would be more convincing with separate results for local attribute edits, object addition/removal, object replacement, geometric edits, background changes, and global appearance edits.

2. Parameter sensitivity and robustness are insufficiently studied.

  • The pipeline contains several consequential choices, including the channel-calibration shrinkage, early-step selection, response normalization, adaptive thresholds, connected-region expansion, safety margin, mask dilation and feathering, harmonic-completion settings, candidate coefficient gamma, and disturbance strength beta. These parameters directly control whether the method preserves or suppresses an edit. It is not clear whether one fixed configuration works robustly across edit categories, image resolutions, mask sizes, backbones, and editing turns, or whether the reported setting was tuned to MTE-Bench.
  • TODO: The paper should provide sensitivity curves, per-category robustness, and results under parameter transfer between MTE-Bench and Banana100. Robustness to random seeds and to deliberately perturbed masks would also be informative.

3. The disturbance is not identifiable from the exterior residual alone.

  • The final-step residual outside the estimated mask is treated as generation-induced degradation. However, an exterior residual may also arise from a legitimate global change requested by the instruction, such as lighting, color tone, artistic style, shadows, reflections, or global scene coherence. It may additionally contain ordinary model prediction error unrelated to the degradation inside the edit region. Consequently, the observed residual does not uniquely identify the nuisance disturbance that should be removed. The current final-output ablations do not directly validate the disturbance estimate.
  • TODO: A stronger evaluation would inject controlled synthetic disturbances with known ground truth and measure estimation accuracy, or compare the estimated field with independently observed reconstruction drift.

4. The spatial smoothness assumption behind harmonic completion is strong.

  • Harmonic completion is appropriate for a low-frequency, spatially smooth field, but many editing artifacts are local or discontinuous. Examples include texture corruption, oversharpening near object boundaries, duplicated structures, facial deformation, high-frequency detail loss, and object-specific geometric changes. Such errors cannot generally be inferred by smoothly extending exterior residuals into the edit region.
  • TODO: The paper should characterize the frequency and spatial structure of the observed degradation, compare harmonic completion with simpler and more expressive alternatives, and report performance as the edit mask becomes larger or the exterior observations become sparse.

5. Localization errors can irreversibly suppress correct edits.

  • The output explicitly restores input pixels outside the estimated support. If the flow-derived mask misses part of the intended edit, this restoration step will overwrite a correct change produced by the backbone. Conversely, an excessively large mask removes the preservation benefit and contaminates the disturbance observations.
  • TODO: The paper does not report mask accuracy against available edit-region annotations, nor does it quantify how final performance changes with false-negative and false-positive mask errors. This is especially important for small objects, thin structures, spatially displaced objects, and edits whose semantic effect extends beyond the most responsive latent locations.

6. The evaluation may favor conservative or incomplete editing.

  • At turns 5 and 10, LPIPS, SSIM, CLIP-I, and DINOv2 measure similarity to the original source image. However, a correctly executed sequence of ten edits may legitimately differ substantially from that image. Pixel restoration and conservative masks can therefore improve most reported fidelity metrics even if the method weakens an intended edit or forgets a previous instruction. CLIP-T for the current instruction does not verify whether all earlier edits are retained.
  • TODO: The benchmark needs a metric or human protocol that jointly checks the cumulative set of requested changes, ideally with per-turn edit success and region-level preservation measured separately.

7. The computational cost is understated.

  • Although disturbance observation reuses a cached prediction, the complete method requires three additional Transformer evaluations per turn according to the paper. In a ten-turn workflow this overhead is substantial. The paper should report end-to-end latency, peak memory, throughput, and total NFEs for every method.
  • TODO: It should also compare against the native editor with an equivalent compute budget, for example by using additional sampling steps or multiple candidates. Without such a comparison, it is unclear whether the improvement comes from the proposed control design rather than simply additional inference computation.
Questions for authors

Please provide your responses on all TODOs in Critical Wealness.

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

Related papers

Discussion (0)

Sign in to comment.