acceptodds
Under review as a conference paper at ICLR 2027

ReState: Noise-Level-Aware State Correction for Training-Free Object Removal in Few-Step Diffusion Models

Abstract

Training-free object removal commonly redirects self-attention away from the target region, but on few-step distilled diffusion models this often leaves object-shaped residuals. We identify a temporal mismatch: a structured prediction forms inside the mask at the first forward, and the short trajectory offers few later chances to revise it, while attention rerouting acts mainly in later steps. Across four distilled backbones, magnitude-matched single-forward interventions show that background-directed state correction affects the final result more than attention rerouting. We therefore propose ReState, which corrects the masked clean prediction toward background evidence retrieved from the input using a noise-level-aware weight that vanishes at the final forward. ReState retains object-excluding attention as a complementary constraint and adds no denoiser evaluations. Across two paired benchmarks and two few-step backbones, ReState achieves the best masked PSNR, SSIM, and LPIPS among training-free methods, improving PSNR by  dB on average over the strongest baseline. On OBER-Test, it also outperforms the 50-step attention-forcing reference on all three metrics at lower latency. The same configuration transfers to two additional distilled backbones without retuning. Code will be released upon publication.

Primary area:
generative models
Keywords:
diffusion models, few-step diffusion, distilled diffusion models, training-free object removal, image editing, image inpainting, state correction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_hdD7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies training-free object removal with few-step distilled diffusion models and proposes ReState. It uses intermediate features from the first denoising pass to match masked positions with similar background positions, then aggregates the corresponding clean latents from the source image to form an explicit background target. This target is used to correct the model’s clean prediction. The correction strength depends on the noise level and becomes zero at the final step, while object-excluding attention remains active throughout sampling. The authors evaluate the method on two paired benchmarks and several distilled backbones. Intervention analyses and ablation studies examine the roles of the background target, correction timing, and attention constraints.

Critical strengths

The method is simple and training-free, works with frozen pretrained models. The experiments are thorough, covering two paired benchmarks and multiple distilled backbones, with ablation studies that support the effectiveness of the main components.

Critical weaknesses

None

Questions for authors
  1. All reported ReState experiments use UNet backbones. Have the authors tried applying it to DiT models such as FLUX or Qwen-Image? If not, what changes would be needed, and what limitations might arise?

  2. Can it also be combined with SARG in Attentive Eraser or adaptive attention suppression in AdaEraser? Have the authors tested whether it provides additional gains over these methods?

3 comments
AuthorsOpenReview ↗

[1/1] Response to Reviewer hdD7

We sincerely thank Reviewer hdD7 for the encouraging review and for recognizing that the method is simple, training-free, and applicable to frozen pretrained models, with thorough experiments and ablations. Below, we answer your two questions. All changes are marked in blue in the revised manuscript.

# [Q1] Extension to DiT Models such as FLUX and Qwen-Image

Yes, we applied ReState to five DiT backbones (FLUX.2-klein-4B and 9B, FLUX 3 Image Klein, Qwen-Image-Flash, and Z-Image-Turbo). On OBER-Test, it raises the masked PSNR from 10.0–11.3 dB to 17.3–19.6 dB at the native number of forwards of each model (Table R1; App. G.2).

  • Required Adaptations. Because ReState is formulated on the clean prediction (App. A.1), three components are adapted: (i) These models predict a velocity under z = (1−σ)x₀ + σϵ, so the clean prediction is x̂₀ = z − σv̂, the weight of Eq. (7) is evaluated at s = σ/(1−σ), and the corrected prediction is mapped back to a velocity. FLUX 3 Image Klein is a flow-map model that also takes the next noise level; its output, the average velocity over the step, is used as v̂ in the same conversion. (ii) Object-excluding attention masks the keys of the object image tokens. (iii) The retrieval features are taken from one transformer block in the first forward. No forward is added.
  • Results (Table R1). ReState improves all four metrics on all five models; on FLUX.2-klein-4B, the masked PSNR rises from 10.89 to 19.44 dB.
  • Limitations. The spatial unit differs: a DiT token packs a 2×2 block of latent positions, whereas our retrieval and correction are defined per latent position, so the corrected latent is re-packed into tokens at every forward. How the correction interacts with the token unit and with flow-matching and flow-map trajectories requires a dedicated analysis, as App. C provides for the UNet backbones; we present the DiT results as an initial extension.

Table R1 (i.e., Tab. G3). ReState on DiT Backbones (OBER-Test). All models use their native resolution and number of forwards (NFE).

BackboneMethodNFEPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
FLUX.2-klein-4BNo Intervention410.890.2610.3620.264
+ ReState (Ours)419.440.5500.1900.118
FLUX.2-klein-9BNo Intervention411.220.2780.3600.259
+ ReState (Ours)419.580.5550.1870.118
FLUX 3 Image KleinNo Intervention1611.260.2900.3550.257
+ ReState (Ours)1618.300.5050.2150.128
Qwen-Image-FlashNo Intervention410.910.2600.4070.283
+ ReState (Ours)417.950.4710.2260.151
Z-Image-TurboNo Intervention810.030.2110.3950.298
+ ReState (Ours)817.300.4510.2460.163

# [Q2] Combination with SARG (Attentive Eraser) and Adaptive Attention Suppression (AdaEraser)

Yes, it can be combined with both, and it provides large gains over them: adding our direct correction raises the re-tuned SARG from 15.34 to 18.57 dB and the adaptive suppression of AdaEraser from 11.71 to 17.79 dB on OBER-Test (Table R2).

  • Why the Combination Is Possible. SARG and adaptive suppression modify attention, whereas the direct correction of ReState is applied to the resulting clean prediction. The attention rule of ReState can thus be replaced by either while the correction is unchanged.
  • Results (Table R2). The correction improves every attention rule, by 3.2 dB for the re-tuned SARG and by 6.1 dB for adaptive suppression. With the official SARG scale of 9, the guidance is too strong for four steps (Sec. 4), and the combination remains low (9.20 dB).
  • Comparison with the Default. None of the combinations exceeds the default ReState (18.75 dB), which uses hard key masking and needs no auxiliary forward, whereas SARG and adaptive suppression each double the denoiser evaluations. We therefore keep hard key masking and report the combinations in App. D.8, with a pointer in Sec. 5.

Table R2 (i.e., Tab. D7). Attention Forcing Without and With the Direct Correction of ReState (OBER-Test, SDXL-Turbo; Masked PSNR ↑ / SSIM ↑ / LPIPS ↓).

Attention Rulew/o Direct Correctionw/ Direct Correction
Hard Key Masking (DesignEdit)15.50 / 0.348 / 0.26918.75 / 0.498 / 0.194
SARG, Re-Tuned Scale (Attentive Eraser)15.34 / 0.335 / 0.27718.57 / 0.495 / 0.201
SARG, Official Scale (Attentive Eraser)7.09 / 0.046 / 0.5149.20 / 0.149 / 0.453
Adaptive Suppression (AdaEraser)11.71 / 0.205 / 0.37217.79 / 0.483 / 0.223

We thank the reviewer again for these two questions. Extending the analysis to further architectures and video diffusion models is an interesting direction for future work.

Thanks for the response and additional experiments. The DiT results and combinations with SARG and AdaEraser answer my two questions. The comparisons with the full official baselines and directly decoded background targets also make the performance gains more convincing. I still see the contribution as incremental, especially since a simple constant-weight schedule performs similarly. However, the method is useful, requires no training, and works well in the tested settings. The remaining limitations, including smooth fills, do not change my recommendation. I maintain my score of 3 (weak acceptance).

AuthorsOpenReview ↗

Thanks to Reviewer hdD7

We sincerely thank Reviewer hdD7 for the follow-up and for the continued positive assessment. We are glad that the DiT results and the combinations with SARG and AdaEraser answered both questions, and that the comparisons with the official baselines and the directly decoded background targets made the gains more convincing.

Your two questions led to the DiT extension (App. G.2) and the combination study (App. D.8) in the revision. We will release our code, evaluation scripts, and per-image outputs to support future work on training-free removal.

Thank you again for the constructive review.

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

This paper studies training-free object removal with few-step distilled diffusion models. The authors observe that attention-based removal methods designed for conventional diffusion sampling transfer poorly to short distilled trajectories: structured object-related content can already form inside the mask at an early denoising forward, while the effect of attention rerouting becomes stronger mainly at later noise levels.

To study this behavior, the paper introduces a matched single-forward intervention analysis that compares object-excluding attention with direct clean-prediction correction from the same latent state and with matched perturbation magnitude. Based on this analysis, the authors propose ReState. ReState retrieves background evidence from unmasked regions of the source image using intermediate denoiser features, directly corrects the masked clean prediction toward this retrieved background proposal, and modulates the correction according to the remaining noise level. Object-excluding attention is retained as a complementary constraint. The method is training-free and reuses features from the first denoiser forward, so it introduces no additional denoiser evaluation.

Experiments are conducted on OBER-Test and RORD using SD-Turbo and SDXL-Turbo, with additional transfer experiments on SDXL-Lightning and Hyper-SDXL.

Critical strengths

1. The intervention analysis is carefully controlled and directly connected to the proposed method.

The matched single-forward experiment is one of the strongest parts of the paper. Attention forcing and direct state correction start from the same latent state at the same noise level, and the direct perturbation is normalized to match the magnitude of the attention-induced change. The additional random and background-orthogonal direction controls further support the interpretation that the improvement depends on moving the state toward meaningful background evidence rather than merely perturbing the trajectory. The analysis is also repeated across four distilled backbones with paired statistical testing.

2. The method is simple and computationally attractive.

ReState uses features already available from the first generation forward, constructs a background proposal from the source image, and applies a closed-form correction to the clean prediction. It therefore does not require removal-specific training, an additional external model, or additional denoiser evaluations. The paper also includes useful ablations on correction strength, timing, evidence quality, retrieval range, feature source, and long-vs-short sampling schedules.

3. The paper makes a meaningful effort to distinguish a few-step regime effect from implementation failure.

The authors additionally evaluate the attention-based baselines in a 50-step setting and show that methods that perform poorly after a direct four-step port can recover substantially at their native long sampling regime. This supports the claim that the difficulty is related to the interaction between the intervention and a short distilled trajectory rather than simply an incorrect baseline implementation.

Critical weaknesses

1. The evaluation protocol and baseline adaptation require further clarification.

The OBER benchmark provides both an object mask and an object-effect mask, but the manuscript only refers to a generic removal mask MM. This distinction is important because ReState anchors all locations outside MM back to the source during sampling, while the reported masked PSNR/SSIM and the masked LPIPS construction exclude errors outside MM. If the object mask is used, shadows, reflections, or contact effects outside the object region cannot be corrected by construction and are also largely excluded from the reported evaluation. If the object-effect mask is used, the method is instead given substantially stronger spatial support. The exact protocol should therefore be stated explicitly, and an effect-aware or whole-image evaluation would help clarify the practical removal quality.

The main few-step comparison fixes the backbone, which is appropriate, but several competing methods were originally designed for substantially longer trajectories. The appendix shows that Attentive Eraser and AdaEraser recover considerably when returned to a 50-step regime. It is therefore important to clarify whether the competing methods were meaningfully retuned for four-step inference, rather than only ported with settings calibrated for their original schedules. There is also a wording issue in Table 2: the caption describes the comparison as using the same four denoiser evaluations, while the effective-forward accounting later reports eight evaluations for Attentive Eraser and AdaEraser.

A separate concern is data provenance. Table B1 checks overlap between the evaluation benchmarks and the removal-specific training/adaptation data of the compared methods, but this does not establish that the frozen foundation models themselves have not seen benchmark imagery during web-scale pretraining. This is particularly relevant for RORD, which predates several of the pretrained backbones used in the paper. The paper should clearly distinguish “no known removal-specific training overlap” from “no foundation-model pretraining overlap.”

2. The diagnosis is meaningful, but the exact novelty and necessity of the proposed state-correction formulation should be positioned more carefully.

The matched intervention analysis provides a useful empirical diagnosis of intervention timing in few-step diffusion. However, direct manipulation of diffusion states or clean predictions is not itself new, as the manuscript correctly acknowledges. The method-level novelty therefore lies primarily in the particular combination of same-backbone background retrieval, explicit background-state correction, and native few-step execution without additional denoiser evaluations.

The related-work positioning should better reflect recent training-free latent-guidance approaches. In particular, Waikit Xiu, Qiang Lu, Junbiao Chen, and Xiying Li, “PredErase: Training-Free Object-and-Effect Removal with Predictive Latent Guidance,” arXiv:2609.00956, 2026 is highly relevant. PredErase is not identical to ReState, but it also uses a frozen generative model, constructs a context-conditioned target for the missing region, and steers the latent trajectory toward that target without removal-specific training. PANDORA, which is already cited in the paper, also combines attention manipulation with localized latent guidance. Thus, the novelty should be framed around the specific retrieved clean-prediction correction and its few-step implementation, rather than around explicit-target or latent/state guidance in general.

I am also not fully convinced that the specific noise-level-aware weighting is sufficiently isolated as an important contribution. With the default configuration, the first three correction weights are already close to strong correction, performance is relatively insensitive to a wide range of λ\lambda, and correcting forwards {2,3}\{2,3\} performs nearly the same as correcting {1,2,3}\{1,2,3\} on PSNR and LPIPS. In contrast, the largest degradation occurs when the final forward is also corrected. A simple constant or binary schedule such as “apply strong correction on all non-final forwards and disable it at the final forward” is therefore an important missing control.

There is also a scope limitation worth clarifying. Because the proposed weight is zero at the final forward, a true one-step backbone would have ω=0\omega=0 at its only forward, making the direct-correction component inactive under the current formulation. Thus, ReState appears to target short-but-multi-step distilled trajectories rather than true one-step removal, and this scope should be stated explicitly.

Questions for authors
  1. Which mask is used on OBER-Test: the released object_mask or object_effect_mask?
    Since ReState anchors all locations outside MM to the source and the reported metrics largely exclude errors outside MM, this choice materially changes the task. If the object mask is used, can the authors additionally report object-effect-region or whole-image evaluation to measure residual shadows/reflections outside the instance mask?

  2. How were the competing attention-based methods adapted to the four-step regime?
    Were their suppression strengths, temporal schedules, guidance parameters, or layer selections retuned for few-step inference? The 50-step experiments show that some methods recover substantially at their native regime, so a comparison against reasonably optimized four-step variants would strengthen the conclusion. Please also clarify the Table 2 statement about “the same four denoiser evaluations,” since the effective-forward accounting reports eight evaluations for Attentive Eraser and AdaEraser.

  3. Can the authors clarify the novelty of ReState relative to recent training-free latent-guidance approaches?
    In particular, please discuss Waikit Xiu, Qiang Lu, Junbiao Chen, and Xiying Li, “PredErase: Training-Free Object-and-Effect Removal with Predictive Latent Guidance,” arXiv:2609.00956, 2026. PANDORA is already cited but also combines attention manipulation with localized latent guidance. What is the precise distinction between these approaches and ReState's explicit background proposal followed by clean-prediction correction?

  4. How important is the specific noise-level-aware weighting in Eq. (7)?
    Can the authors compare it with simple constant or binary schedules, e.g., applying the same strong correction to all non-final forwards and setting the final-forward correction to zero? The existing λ\lambda and timing ablations suggest that avoiding correction at the final forward may be more important than the exact continuous weighting function.

  5. How should the first-forward diagnosis be interpreted relative to the deployed method?
    The intervention analysis shows a clear first-forward advantage for direct correction, while the deployment ablation shows that correcting {2,3}\{2,3\} is almost indistinguishable from correcting {1,2,3}\{1,2,3\} on PSNR and LPIPS. Is the first-forward experiment primarily evidence for the general intervention mechanism rather than evidence that first-forward correction is itself necessary?

  6. Can the authors clarify possible benchmark overlap with the pretraining data of the frozen foundation models?
    Table B1 addresses the removal-specific training/adaptation data of the compared methods, but not the web-scale pretraining data of SDXL, SD-Turbo, and related backbones. In particular, RORD predates several of these models. Can direct or near-duplicate benchmark overlap be ruled out or assessed? If not, the manuscript should explicitly limit its claim to the absence of known removal-specific training overlap.

  7. What is the intended lower-bound scope of “few-step” ReState?
    Under the current formulation, the correction weight vanishes at the final forward. For a true one-step backbone, the only forward is also the final forward, so the direct-correction term would be zero. Is ReState therefore intended specifically for short multi-step trajectories rather than one-step distilled models?

5 comments
AuthorsOpenReview ↗

[1/3] Response to Reviewer CJpo

We sincerely thank Reviewer CJpo for the thorough and constructive review. We appreciate your recognition that the intervention analysis is carefully controlled and directly connected to the method, that the method is simple and computationally attractive, and that the 50-step experiments distinguish a few-step regime effect from an implementation failure. Below, we respond point by point, with each weakness grouped with its related question. All changes are marked in blue in the revised manuscript.

# [W1/Q1] Mask Protocol, Object-Effect Masks, and Whole-Image Evaluation

The submission uses the released object mask for all 163 images of OBER-Test. As you describe, effects outside this mask are preserved by the anchoring and are largely excluded from the masked metrics. With the object-effect mask as input, the unchanged ReState improves the effect-only region by 6.5 dB and the whole image by 1.5 dB (Table R1).

  • Protocol (Revised App. B). App. B now states which mask is used for generation and for evaluation on each benchmark.
  • Object-Effect Masks and Whole-Image Evaluation on OBER-Test (Table R1). For each method, we generate once with the object mask and once with the object-effect mask, without re-tuning, and score both outputs on the object region, the effect-only region, and the whole image. ReState is the best method in every region under both masks; its whole-image PSNR is 24.49 dB with the object mask and 25.95 dB with the object-effect mask.
  • RORD. RORD annotates each object with a coarse mask that also covers its effects. For this experiment, we use RORD-Val, the 343 validation frames for which the authors of OBER additionally released refined object masks. On this set, the effect region is much larger (a median of 11.0% of the frame vs. 2.5% on OBER-Test), and many of the masks are coarse polygons that also cover unaffected background regions (Fig. F1): in the median image, 58% of the effect-only pixels already match the ground truth within 8/255 in the input, against 9% on OBER-Test. Regenerating this region lowers the effect-only and whole-image scores of all five methods. ReState remains the best method under both masks. The complete tables are Tab. F3 and Tab. F4 (App. F.3).

Table R1 (i.e., Tab. F3, Selected Rows). Object Mask vs. Object-Effect Mask as Input on OBER-Test (SDXL-Turbo). DesignEdit is the strongest 4-step baseline.

MethodInput MaskObject Region PSNR ↑Effect-Only Region PSNR ↑Whole-Image PSNR ↑Whole-Image LPIPS ↓
+ DesignEdit (AAAI'25)Object15.5014.9023.240.139
Object-Effect16.3118.6224.060.143
+ ReState (Ours)Object18.7515.5924.490.125
Object-Effect19.6822.1125.950.123

# [W1/Q2] Adaptation of the Competing Methods to the Four-Step Regime

The ports of the submission apply each method's attention rule with its official hyperparameters inside the shared four-step pipeline. We add a re-tuned port and the official implementations at their native step counts: re-tuning raises Attentive Eraser from 7.09 to 15.34 dB, the official DesignEdit and Attentive Eraser reach 18.51 and 18.19 dB, and 4-step ReState reaches 18.75 dB on OBER-Test (Tab. 5; Tab. E4).

  • What the Ports Keep (Tab. B3). A new table lists every component of the official implementations and whether the port keeps it. The DesignEdit port keeps only hard key-masking, without inversion and guidance. The Attentive Eraser port keeps the official scales and applies its attention redirection to all self-attention layers, whereas the official code restricts it to the up-block layers. App. B.3 now states these differences for each method.
  • Re-Tuning. We re-tune the guidance scale of Attentive Eraser, the parameter behind its failure at four steps, on each Turbo backbone; its layer selection and step schedule are kept as in the port. ReState remains 3.4 dB (SDXL-Turbo) and 3.2 dB (SD-Turbo) higher (Tab. B4).
  • Official Implementations. We also run the public code of DesignEdit (50 inversion + 50 sampling steps) and Attentive Eraser (40 steps) with all components active. ReState is higher than both on PSNR, SSIM, and LPIPS, while being 14.2× and 12.4× faster, and it is also higher than both on RORD and MULAN (Tab. E4). These results are now in the main text (Tab. 5).
  • Caption of Tab. 2. Attentive Eraser and AdaEraser use eight evaluations, as Tab. 5 reports. To match this, the caption now reads "the same four-step sampling schedule".

(Continued in Next Response)

AuthorsOpenReview ↗

[2/3] Response to Reviewer CJpo

# [W2/Q3] Novelty Relative to PredErase and PANDORA

Thank you for pointing us to PredErase. We agree that the novelty lies in the retrieved clean-prediction correction and its few-step implementation, not in explicit targets or latent guidance in general; Sec. 2 and Tab. 1 are revised accordingly. The three methods differ in how the target is obtained and how it is incorporated into sampling: ReState is the only one that builds the target from the same forward of the same backbone, without an external model, and applies it in closed form without any additional evaluation (Table R2).

  • PANDORA. PANDORA runs Stable Diffusion for 50 steps after DDIM inversion. Its guidance (their Eq. (6), rearranged) acts on the noise prediction inside the mask:

    ε̂ = ε_u + α (ε_c − ε_u),

    where ε_c is the prediction with its attention manipulation and ε_u is the reference prediction without it. This is the same form as SARG in Attentive Eraser (their Eq. (18)):

    ε̂ = ε + s (ε_AAS − ε),

    where ε_AAS is the prediction with the redirected self-attention. Both extrapolate from a reference noise prediction toward an attention-modified one, and neither defines a target for the masked region nor moves the latent toward one. The guidance of PANDORA is thus an attention-based guidance on the noise prediction, combined with its own attention manipulation and inversion-based reference. Unlike ReState, it does not construct a clean-latent target B.

  • PredErase. PredErase runs FLUX.2 for 14 steps and obtains its target from an external model: a frozen I-JEPA predictor infers a representation of the masked region from its context, and an alignment loss is back-propagated through the decoder to update the latent at selected steps. As you note, it is the closest in spirit.

  • ReState. ReState needs no external prior. Its target B is a clean latent assembled from unmasked source positions with the features of the first forward, so it can be written into the clean prediction directly (Eq. (7)), with no back-propagation and no additional forward. This makes it applicable within a four-step budget, a regime in which neither of the two methods is evaluated.

Table R2 (i.e., Tab. 1, Extended). Comparison with Training-Free Methods That Guide the Masked Region.

PANDORA (ICME'26)PredErase (arXiv'26)ReState (Ours)
External ModelNoneI-JEPANone
Target for the Masked RegionNoneRepresentation predicted by I-JEPAClean latent B retrieved with the features of the first forward
How It Guides SamplingGuidance on the noise prediction (same form as SARG)Gradient through the decoder at selected stepsClosed-form correction of the clean prediction
Additional EvaluationsOne per stepDecoder and I-JEPA passes with back-propagationNone

# [W2/Q4] Importance of the Noise-Level-Aware Weighting in Eq. (7)

Releasing the final forward is indeed the most important property (App. D.2), and we add the constant and binary schedules that you suggest. A strong constant weight with a released final forward is close to Eq. (7) (18.68 to 18.76 dB vs. 18.75 dB; Table R3), a weak one loses 1.2 dB, and keeping the correction at the final forward loses up to 1.5 dB and raises LPIPS from 0.194 to 0.284 (App. D.2).

  • Results (Table R3). Constant weights between 0.794 and 1.0, where 1.0 is the binary schedule (1, 1, 1, 0), are within 0.07 dB of the default. A constant weight of 0.5 loses 0.2 dB, and 0.25 loses 1.2 dB. With the first three weights of Eq. (7) fixed, raising the final weight to 0.2, 0.5, and 0.794 degrades LPIPS monotonically, from 0.194 to 0.210, 0.251, and 0.283 (App. D.2).
  • What Eq. (7) Contributes. The results support two structural properties, a strong correction while noise remains and none at the final forward, and not the exact intermediate weights. Eq. (7) provides both in closed form from the noise level, so no constant or window has to be chosen, and the same rule gives the weights of shorter schedules (see [W2/Q7]).
  • Revision. Tab. D2 now includes the constant schedules, and Sec. 3.4 states the role of the weighting as above.

Table R3 (i.e., Tab. D2, Constant Rows). Correction Weights at the Four Forwards on OBER-Test (SDXL-Turbo). Only the weights differ between rows; every schedule releases the final forward.

ScheduleWeights (ω₁, ω₂, ω₃, ω₄)PSNR ↑LPIPS ↓
Eq. (7) (Default)(0.993, 0.954, 0.794, 0)18.750.194
Constant 0.25(0.25, 0.25, 0.25, 0)17.540.217
Constant 0.5(0.5, 0.5, 0.5, 0)18.530.197
Constant 0.794(0.794, 0.794, 0.794, 0)18.760.194
Constant 0.914 (Mean of Default)(0.914, 0.914, 0.914, 0)18.730.195
Binary(1, 1, 1, 0)18.680.196

(Continued in Next Response)

AuthorsOpenReview ↗

[3/3] Response to Reviewer CJpo

# [Q5] Interpretation of the First-Forward Diagnosis

Your reading is correct: the first-forward experiment is evidence for the general intervention mechanism, namely which kind of intervention reaches the final image, and not evidence that the first-forward correction itself is necessary.

  • In the Deployed Method. When the second and third forwards are corrected, the first-forward correction is redundant, as App. D.2 reports (18.74 vs. 18.75 dB). The later corrections overlap with it.
  • When It Matters. The first-forward correction carries the gain when later corrections are not available. With a two-step schedule, the correction is applied only at the first forward, and ReState reaches 16.38 dB against 13.12 dB for attention forcing alone (Tab. D6).
  • Revision. Sec. 5 and App. D.2 now state this distinction between the diagnosis and the deployed schedule explicitly.

# [W1/Q6] Benchmark Overlap with the Pretraining Data of the Backbones

We agree that the two claims should be stated separately. OBER-Test was captured by its authors and first released after the four distilled backbones, so its images could not have been part of their pretraining data. For RORD, which predates them, this cannot be established, and the revision states the claim as the absence of known removal-specific training overlap.

  • OBER-Test. Its image pairs were captured by the authors of OBER with fixed cameras for the dataset and first released in 2025, after the four distilled backbones of our paper (SDXL-Turbo and SD-Turbo in 2023; SDXL-Lightning and Hyper-SDXL in 2024).
  • RORD. RORD was released in 2022 and predates these backbones. Its frames come from scenes recorded for the dataset and are distributed as a dataset archive, not as web image–text pairs, but the pretraining sets are not fully documented (SDXL reports only an internal dataset), so we cannot rule out overlap.
  • Effect on the Comparison. All methods within each block of Tab. 2 use the same frozen backbone, so the comparison does not depend on differences between pretrained models.
  • Revision. App. B (Tab. B1) now distinguishes "no known removal-specific training overlap" from "no foundation-model pretraining overlap" and states the above for each benchmark.

# [W2/Q7] Lower-Bound Scope of "Few-Step" ReState

Your understanding is correct: on a one-step model, the weight of the only forward is zero, the direct correction is inactive, and the output of ReState is identical to that of attention forcing alone. ReState targets short multi-step trajectories with at least two forwards.

  • Step Count (App. D.7; Tab. D6). We run SDXL-Turbo and SD-Turbo with one to four steps. On OBER-Test with SDXL-Turbo, ReState reaches 16.38, 18.27, and 18.75 dB with two, three, and four steps, 3.2 to 3.5 dB above attention forcing alone at the same step count. With one step, the two outputs are identical.
  • Revision. Sec. 3 and Sec. 5 now state that ReState addresses few-step trajectories with two or more forwards and not one-step models. One-step removal, where no later forward can refine a correction, is left as future work.

We hope that the effect-aware evaluation, the re-tuned and official baselines, the positioning against PredErase and PANDORA, and the schedule and step-count controls address your concerns. If any point remains unclear, or if you would like to see further experiments, please let us know; we will respond and add the results as quickly as possible.

Reviewer_1ohSOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses training-free object removal on few-step distilled diffusion models (SDXL-Turbo, SD-Turbo, with transfer to SDXL-Lightning and Hyper-SDXL, all at 4 steps). It argues that attention-based removal ("attention forcing", i.e., masking object keys in self-attention) leaves object-shaped residuals in this regime because the short trajectory commits to a structured masked prediction at the first forward. A matched single-forward intervention analysis is used to support this. The proposed method, ReState, (i) retrieves a background proposal B for each masked latent position by matching first-forward UNet features to unmasked positions and averaging the top-1% clean source latents, (ii) corrects the masked clean prediction toward B with a weight ω_t = λσ²_{t−1}/(1+λσ²_{t−1}), derived as the minimizer of a proximal objective, which vanishes at the final forward, and (iii) keeps hard object-key masking and out-of-mask latent anchoring throughout. It adds no denoiser evaluations. On OBER-Test and RORD, ReState reports the best masked PSNR/SSIM/LPIPS among training-free methods (+3.17 dB masked PSNR over DesignEdit on average) and matches or exceeds 50-step attention-forcing configurations at 6.7× lower latency.

Critical strengths
  1. A practical and well-motivated problem, addressed at no extra inference cost. Training-free removal on few-step distilled models is useful, and the paper clearly shows that attention-based removal designed for 50-step sampling degrades in this regime. ReState is simple, needs no training, reuses the first forward for retrieval, and stays at 4 NFEs. If the gains hold under a fair comparison, this is a practically relevant contribution.

  2. Extensive analysis and ablations. The paper ablates the correction strength, the correction window, the retrieval ratio, the feature source, and the evidence source (including an oracle B), and it uses paired permutation tests throughout. The finding that a first-forward perturbation is preserved on the 4-step trajectory but attenuated on the 50-step one (App. C.4) is informative for the few-step regime.

Critical weaknesses
  1. Baselines are weakened and ported to the 4-step setting. All training-free baselines are re-implemented in the authors' own 4-step pipeline. DesignEdit loses its DDIM inversion and is reduced to hard key-masking (App. B), and Attentive Eraser and AdaEraser keep their 50-step hyperparameters without re-tuning. A method's components are part of the method and leak no information, so removing them is not a fair way to control variables. The port strongly affects the results. Attentive Eraser drops below the no-intervention baseline at 4 steps (7.09 vs 9.86 dB), but reaches 18.79 dB at 50 steps (Table E3), on par with ReState (18.75 dB). The main table should report both the 4-step ports and each baseline's official version at its native step count, together with latency.

  2. The necessity of the first two forwards is unclear. With ω ≈ 0.993/0.954/0.794/0, the masked prediction is almost fully replaced by B in the first two forwards. Table D2 shows that correcting the first forward is redundant ({2,3} ≈ full). The {3} setting in Table D2 still runs the first two forwards without correction, and at the full 4-NFE cost. A simpler alternative would be to retrieve B in the first forward, insert B plus noise directly at the 2nd or 3rd step, and run only the remaining forwards to add details. This would be faster and may perform similarly, which would make the per-step correction unnecessary.

  3. Results are blurry, and the metrics favor blurry fills. The masked regions in Fig. 3/E1/E2 are visibly smooth, consistent with B being an average of retrieved latents and with App. C.5 (high-band energy 0.51× of GT). Masked PSNR/SSIM are lenient toward smooth fills: a global background mean used as B already reaches 18.16 dB (Table D3), above all training-free baselines. Perceptual metrics or a user study are needed, and the figures should be shown without contour overlays.

  4. Related work on few-step training-free editing is missing. The paper claims that prior work on few-step distilled models mainly studies generation quality and speed. However, training-free editing on distilled backbones has been studied by TurboEdit, InfEdit, and ChordEdit. ChordEdit in particular starts from the same observation that training-free editors fail in the few/one-step regime. These works should be discussed.

  5. The datasets do not match the method. Both OBER and RORD are object-and-effect removal datasets whose ground truths also remove shadows and reflections, and they provide object-effect masks. ReState uses object-only masks and does not remove effects, and the masked metrics never measure the region outside the mask, so remaining shadows are not penalized. The authors should add object-only removal benchmarks evaluated with full-image metrics, and take the object-effect masks into account.

Minor issues (typos):

Sec. 4: broken reference "Appendix ??".

Table 1 caption: duplicated punctuation ";;".

Questions for authors
  1. Can you add the official versions of the baselines at their native step counts to the main table, next to the 4-step ports, with latency?
  2. How does the simpler variant (retrieve B in the first forward, insert B plus noise at the 2nd or 3rd step, and 3. run only the remaining forwards) compare in quality and speed?
  3. Can you report perceptual metrics or a user study, and show figures without contour overlays?
  4. Can you report results on object-only removal benchmarks with full-image metrics, and on OBER/RORD using the object-effect masks?
6 comments
AuthorsOpenReview ↗

[1/4] Response to Reviewer 1ohS

We sincerely thank Reviewer 1ohS for the detailed and constructive review. We appreciate your recognition of the practical and well-motivated problem, of the fact that ReState needs no additional denoiser evaluations, and of the extensive analysis and ablations, including the finding that a first-forward perturbation is preserved on the four-step trajectory but attenuated on the 50-step one (App. C.4). Below, we respond point by point, with related weaknesses and questions grouped under one subtitle. All changes are marked in blue in the revised manuscript.

# [W1/Q1] Baselines: Official Implementations at Their Native Step Counts and Latency

Thank you for this suggestion. We ran DesignEdit and Attentive Eraser with their official implementations at their native step counts, and evaluated our implementation of AdaEraser at 50 steps. On OBER-Test, DesignEdit (50 inversion + 50 sampling steps) and Attentive Eraser (40 steps) reach 18.51 and 18.19 dB. 4-step ReState reaches 18.75 dB: it is higher than both on PSNR, SSIM, LPIPS, and DISTS, while being 14.2× and 12.4× faster (Table R1). It is also higher than both on RORD and MULAN.

  • Official Implementations (Table R1). We run the public code of DesignEdit and Attentive Eraser through a wrapper, without editing their files, at 512×512. All components of the official procedures are active: for DesignEdit, DDIM inversion, the two-layer generation with classifier-free guidance 7.5, layer fusion, and proximal guidance; for Attentive Eraser, 40 of 50 DDIM steps (strength 0.8), SARG with scale 9, and similarity suppression. The latency is end-to-end, including inversion, and measured under the same hardware and software environment as ReState.
  • AdaEraser. AdaEraser has no public implementation, so we implement it from its paper (Sec. 3.3, Eqs. (2) to (7)). The paper does not state a step count, and we report 50 steps on SDXL 1.0 (15.66 dB). The AdaEraser paper itself reports that, on 4-step SDXL-Lightning, the method shows "a clear degradation in removal and reconstruction quality" and "does not directly transfer well to highly accelerated distilled models" (their App. D), which is consistent with its 4-step row in Tab. 2.
  • 4-Step Ports and Re-Tuning (App. B.3). As you point out, the DesignEdit port is reduced to hard key-masking and coincides with the attention-forcing-only row of Tab. 4; the revision labels it accordingly. The Attentive Eraser port keeps the official scales and applies its attention redirection to all self-attention layers, whereas the official code restricts it to the up-block layers. App. B now states these differences, and Tab. B3 lists which components of each official implementation the port keeps. Re-tuning the SARG scale of the Attentive Eraser port raises it from 7.09 to 15.34 dB (Tab. B4), and ReState remains 3.4 dB higher.
  • Revision. Tab. 2 keeps the controlled comparison on the same few-step backbone. The official results are now in the main text (Tab. 5, with their latency), and the abstract and Sec. 4 compare against them; App. E.3 reports them on all three benchmarks (Tab. E4). Tab. E3 applies each attention rule inside our shared pipeline, so its rows are now named by the attention rule (hard key masking, SARG, and adaptive suppression) and not by the method, and the factor of 6.7× is stated only for 50-step hard key masking in that pipeline.

Table R1 (i.e., Tab. 5, Tab. E4, and Tab. E3). Baselines in Their Official and Ported Versions. The first five metric columns are on OBER-Test; the last two are the masked PSNR on RORD and MULAN. Latency is end-to-end and relative to ReState; our AdaEraser implementation calls its two branches sequentially. The latency of the 4-step ports is reported in Tab. 2.

MethodVersionStepsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓Latency ↓RORD PSNR ↑MULAN PSNR ↑
DesignEdit (AAAI'25)Official Implementation, SDXL50 Inversion + 5018.510.4830.2060.14014.2×16.5215.85
Attentive Eraser (AAAI'25)Official Implementation, SDXL4018.190.4460.2450.20412.4×15.7115.17
AdaEraser (ICML'26)Our Implementation, SDXL5015.660.3990.2740.18623.9×12.0812.44
DesignEdit4-Step Port, SDXL-Turbo415.500.3480.2690.199–14.3414.07
Attentive Eraser4-Step Port, SDXL-Turbo47.090.0460.5140.387–7.357.21
Attentive Eraser4-Step Port, Re-Tuned415.340.3350.2770.209–14.1814.00
AdaEraser4-Step Port, SDXL-Turbo411.710.2050.3720.292–9.4710.15
ReState (Ours)SDXL-Turbo418.750.4980.1940.1291.0×17.2116.63

(Continued in Next Response)

AuthorsOpenReview ↗

[2/4] Response to Reviewer 1ohS

# [W2/Q2] One-Shot Insertion of B vs. Per-Forward Correction

App. D.4 covers the earliest entry point, and we extend it to every entry point. The result depends strongly on the noise level at which B is inserted: on OBER-Test with SDXL-Turbo, inserting B before the 2nd, 3rd, and 4th forward gives 17.90, 18.53, and 18.61 dB, against 18.75 dB for the per-forward correction (Table R2).

  • Setup. All variants share the first forward (needed to retrieve B), the retrieved B, object-excluding attention, background anchoring, and the noise. At the entry noise level σ, the masked region is re-initialized as B + σϵ, and the forwards before the entry are skipped.
  • Results (Table R2). Inserted at the 2nd step, B loses 0.85 dB at the same 4 NFEs, since the following forwards re-predict the masked content without any correction. Inserted later, B is preserved, and the result comes within 0.14 to 0.22 dB of ReState with 2 or 3 NFEs. The 2-NFE variant is lower than ReState in all eight benchmark–backbone settings of the paper, by 0.06 to 0.22 dB (App. D.4).
  • Is the Per-Forward Correction Needed? A late one-shot insertion is a useful faster option, since it keeps most of the gain with fewer forwards. The per-forward correction remains our default for two reasons. First, it gives the highest PSNR in Table R2, 0.14 to 0.85 dB above the one-shot variants. Second, it needs no choice of an entry step: the one-shot variant is effective only when B is inserted at a suitable noise level, whereas ReState applies B at every forward with a weight set by the noise level (Eq. (7)). The one-shot results thus follow the same noise-level-aware principle.
  • Revision. App. D.4 now reports all entry points, and the 2-NFE insertion as a faster variant (0.70× the end-to-end time).

Table R2 (i.e., Tab. D4, Bottom). One-Shot Insertion of B at Different Entry Points vs. Per-Forward Correction on OBER-Test.

VariantEntry Noise Level σNFEPSNR ↑SSIM ↑LPIPS ↓
Insertion Before the 2nd Forward (App. D.4)4.082417.900.4540.211
Insertion Before the 3rd Forward1.613318.530.4880.196
Insertion Before the 4th Forward0.693218.610.4980.199
ReState (Per-Forward Correction)–418.750.4980.194

# [W3/Q3] Blurry Fills, Perceptual Metrics, and Figures Without Contour Overlays

We agree with this concern and measured it directly. A ground truth whose masked region is blurred with a Gaussian kernel (σ = 8) reaches 24.07 dB in masked PSNR, above every method, so masked PSNR and SSIM do reward smooth fills. We therefore add DISTS, CMMD, and ReMOVE to LPIPS, and ReState is the best training-free method on all four, on both backbones (Table R3). The figures added in the revision (Figs. F1, G1, G2, and H3) show the outputs without contour overlays.

  • Blur Control (First Row of Table R3). Scored as if it were a removal result, the blurred ground truth is better than every method in masked PSNR and SSIM but worse than ReState in LPIPS, CMMD, and ReMOVE, so these metrics do not reward smoothness in the same way.
  • Perceptual and Distribution Metrics (Table R3). Relative to the strongest 4-step baseline (DesignEdit), ReState lowers LPIPS from 0.269 to 0.194, DISTS from 0.199 to 0.129, and CMMD from 1.582 to 1.358, and raises ReMOVE from 0.849 to 0.929. Tab. F1 reports both backbones.
  • A Smooth Fill Alone. The global-mean row of Tab. D3 uses the mean as B inside the same correction, so it is not a plain smooth fill. Decoded directly without denoising, the global mean and the retrieved B give 15.13 and 16.55 dB, against 18.16 and 18.75 dB when the same B is used in the per-forward correction (Tab. 6).
  • Remaining Limitation. The fill of ReState is still smoother than the real background (0.51× the high-band energy of the ground truth; App. C.5), which we state as a limitation.

Table R3 (i.e., Tab. F1, SDXL-Turbo Rows). Additional Metrics on OBER-Test. The first row is the blur control (σ = 8), reported in the text of App. F.1.

MethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓CMMD ↓ReMOVE ↑
Ground Truth, Blurred (σ = 8)24.070.6040.2210.1351.6390.656
SDXL-Turbo (4 Steps)9.860.1330.4350.3181.9130.678
+ Attentive Eraser (AAAI'25)7.090.0460.5140.3872.7470.684
+ Attentive Eraser, Re-Tuned15.340.3350.2770.2091.6550.833
+ DesignEdit (AAAI'25)15.500.3480.2690.1991.5820.849
+ AdaEraser (ICML'26)11.710.2050.3720.2921.7220.709
+ ReState (Ours)18.750.4980.1940.1291.3580.929

(Continued in Next Response)

AuthorsOpenReview ↗

[3/4] Response to Reviewer 1ohS

# [W4] Training-Free Editing on Few-Step Models: TurboEdit, InfEdit, and ChordEdit

Thank you for pointing out these works. We now discuss TurboEdit, InfEdit, and ChordEdit in Sec. 2 and evaluate all three on OBER-Test: inside the mask, they stay at the level of the output without any intervention (9.29 to 9.95 dB vs. 9.86 dB), whereas ReState reaches 18.75 dB (Table R4).

  • What These Methods Study. All three edit an image from a source prompt to a target prompt on a few-step model: TurboEdit (SIGGRAPH Asia'24) with an inversion adapted to SDXL-Turbo, InfEdit (CVPR'24) with an inversion-free consistency sampler, and ChordEdit (CVPR'26) with a one-step edit direction averaged over a time window. As you note, ChordEdit starts from the same observation as our paper, that training-free editors break down in the few-step regime, and Sec. 2 now states this shared starting point.
  • Setup and Results (Table R4). These editors take a pair of prompts and no mask. For each image, a vision-language model (Qwen2.5-VL-7B, fixed templates, greedy decoding) names the object inside the mask and writes a source caption with the object and a target caption without it. Each method runs its official code with its default backbone and settings. The object largely remains in the image: the masked PSNR is 9.29 to 9.95 dB, and DISTS is 0.290 to 0.304 vs. 0.129 for ReState.
  • Editing Generalists vs. a Removal Specialist. To give an editor the same spatial information as ReState, we also run InfEdit with the ground-truth mask in its local blending. Its background is then preserved as well as with ReState (27.84 vs. 27.59 dB), but its masked PSNR barely changes (9.29 to 9.69 dB). Knowing where to edit is therefore not enough: removal also needs a target for what lies behind the object, and a prompt that omits the object does not provide one. Text-driven editors cover many edit types with a prompt alone, whereas ReState covers only mask-based removal and, in return, builds this background target from the image itself and writes it into the clean prediction. We see the two lines as complementary (App. G.1).

Table R4 (i.e., Tab. G1). Few-Step Text-Driven Editors on Object Removal (OBER-Test).

MethodBackbonePSNR ↑SSIM ↑LPIPS ↓DISTS ↓Background PSNR ↑
TurboEdit (SIGGRAPH Asia'24)SDXL-Turbo, 4 Steps9.950.2250.4200.29018.72
InfEdit (CVPR'24)LCM-Dreamshaper-v7, 12 Steps9.290.1460.4370.30417.71
InfEdit + Ground-Truth Mask (Our Variant)LCM-Dreamshaper-v7, 12 Steps9.690.1620.4270.29827.84
ChordEdit (CVPR'26)SD-Turbo, 1 Step9.890.1900.4180.29118.42
ReState (Ours)SDXL-Turbo, 4 Steps18.750.4980.1940.12927.59

(Continued in Next Response)

Reviewer_1ujDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies training-free object removal on few-step distilled diffusion models such as SDXL-Turbo. Earlier training-free methods stop the model from looking at the object while it repaints that area. The authors find this doesn’t work well for few-step distilled diffusion models.

The model mostly decides what to put in the hole in the first step, but this trick has little effect until the last steps, when it is too late to change much. So, for each spot in the hole, the authors find background patches elsewhere in the image that look similar and average them. Over the first three steps, they push the model’s guess toward that average, then leave the last step alone so the model can fill in detail. They also keep the view-blocking trick.

For evaluation, on two datasets and two few-step backbones, it gets better PSNR scores than the best earlier training-free method, and the same settings also work on other models.

Critical strengths
  1. The method is simple and cheap. It's training free, and the hyperparameters works on other backbones. That might have some applications in mobile computing.

  2. The authors report the global-mean baseline, the oracle, and the 50-step transfer that doesn’t work, which shows where gains come from.

Critical weaknesses
  1. Most of the gain isn’t from the paper’s idea. global background mean already have good results, and w/o the first-step correction changes little, which is not in favor of your story.

  2. The baselines are weak. DesignEdit is the paper’s own ablation, Attentive Eraser runs with a guidance scale that breaks Turbo models, and there is no comparison to LaMa.

  3. I kind of concern that if this is enough for acceptance since without the story, the method is patch filling and a known data-consistency step.

Questions for authors

If you skip the diffusion model entirely and just decode B into the hole, how close does that get to ReState?

8 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer 1ujD

We sincerely thank Reviewer 1ujD for the careful review and for recognizing that our method is simple, lightweight, and training-free with hyperparameters that transfer to other backbones, and that the global-mean, oracle, and 50-step transfer results show where the gains come from. Below, we respond point by point. All changes are marked in blue in the revised manuscript.

# [W1/Q1] Where the Gain Comes From: Decoding B Directly vs. Per-Forward Correction

We ran this test. If we skip the denoising and decode B into the mask, the result is 16.55 dB with the retrieved B and 15.13 dB with the global background mean. Using the same B within ReState gives 18.75 and 18.16 dB (Table R1), which is 2.2 to 3.0 dB above the direct decode, with better perceptual metrics as well.

  • The Global-Mean Row Is Not a Mean Fill. The global-mean row of Tab. D3 (18.16 dB) uses the mean as B inside the same per-forward correction, so it already incorporates our proposed per-forward correction; the mean fill itself reaches 15.13 dB. With the retrieved B, the direct decode is flat and blurry (LPIPS 0.299, CMMD 2.152), and ReState improves it to 0.194 and 1.358. Most of the gain over a plain fill thus comes from how B is used during sampling. These results are now in the main text (Tab. 6).
  • Removing the First-Forward Correction. As App. D.2 reports, correcting only the second and third forwards gives nearly the same result as the full schedule (18.74 vs. 18.75 dB). The first-forward analysis of Tab. 3 is a diagnostic of which kind of intervention reaches the final image, not a claim that the first forward must be corrected. It carries the gain when later corrections are absent: with a two-step schedule, ReState reaches 16.38 dB against 13.12 dB for attention forcing alone (Tab. D6).

Table R1 (i.e., Tab. 6 and Tab. D3). Evidence B Decoded Directly vs. Used in the Per-Forward Correction (OBER-Test, SDXL-Turbo). "Decoded Directly" replaces the masked region of the clean source latent with B and decodes it without denoising.

Evidence BUse of BPSNR ↑SSIM ↑LPIPS ↓DISTS ↓CMMD ↓
Global MeanDecoded Directly15.130.4660.3370.2032.112
Global MeanPer-Forward Correction18.160.4810.2140.1331.398
RetrievedDecoded Directly16.550.4480.2990.1962.152
RetrievedPer-Forward Correction (ReState)18.750.4980.1940.1291.358

# [W2] Baselines: Official Implementations, Re-Tuned Attentive Eraser, and LaMa

We added the three comparisons that you point to. With their official implementations and native step counts, DesignEdit (50 inversion + 50 sampling steps) and Attentive Eraser (40 steps) reach 18.51 and 18.19 dB on OBER-Test, and LaMa reaches 17.49 dB; 4-step ReState reaches 18.75 dB (Table R2).

  • DesignEdit and Attentive Eraser in Their Official Versions. We run the public code of both methods in full (DDIM inversion and classifier-free guidance for DesignEdit; SARG at its official scale for Attentive Eraser). ReState is higher than both on PSNR, SSIM, and LPIPS, while being 14.2× and 12.4× faster, and it is also higher on RORD and MULAN (Tab. 5; Tab. E4).
  • The 4-Step Ports. As you point out, the DesignEdit port is reduced to hard key-masking and coincides with the attention-forcing-only row of Tab. 4; the revision labels it accordingly. Re-tuning the guidance scale of Attentive Eraser raises it from 7.09 to 15.34 dB (App. B.3), and ReState remains 3.4 dB higher.
  • LaMa. We add LaMa, a network trained for inpainting, to Tab. E1. ReState is higher in masked PSNR on two of the three benchmarks: by 1.3 dB on OBER-Test and by 1.2 dB on the object-only MULAN benchmark, where it is also higher in SSIM (0.457 vs. 0.380) at a similar LPIPS (0.268 vs. 0.272; App. F.2). LaMa is better in SSIM and LPIPS on OBER-Test (0.540 and 0.169 vs. 0.498 and 0.194) and on all three metrics on RORD (17.30 dB, 0.494, and 0.210 vs. 17.21 dB, 0.451, and 0.273). Without any training, ReState thus performs on par with a trained inpainting network.

Table R2 (i.e., Tab. 5, Tab. E4, and Tab. E1, Selected Rows). Stronger Baselines. PSNR, SSIM, and LPIPS are on OBER-Test; the last two columns are the masked PSNR on RORD and MULAN.

MethodVersionStepsPSNR ↑SSIM ↑LPIPS ↓Latency ↓RORD PSNR ↑MULAN PSNR ↑
DesignEdit (AAAI'25)Official, SDXL50 + 5018.510.4830.20614.2×16.5215.85
Attentive Eraser (AAAI'25)Official, SDXL4018.190.4460.24512.4×15.7115.17
LaMaTrained Inpainting Network–17.490.5400.1690.14x17.3015.46
ReState (Ours)SDXL-Turbo418.750.4980.1941.0×17.2116.63

(Continued in Next Response)

AuthorsOpenReview ↗

[2/2] Response to Reviewer 1ujD

# [W3] Contribution Beyond Patch Filling and Data Consistency

We understand this concern. Neither filling from background patches nor modifying a clean prediction is new by itself, as Sec. 2 states; the revision adds the classical references for both (Sec. 2 and Sec. 3.3). The contribution is the finding of which intervention reaches the final image on a few-step trajectory, and a correction built on it, which writes retrieved background evidence into the clean prediction while noise remains and releases the final forward. The new results show that both parts matter: patch filling alone gives 16.55 dB, and a correction that is too weak or kept at the final forward loses 1.2 to 1.5 dB.

  • Finding. Existing training-free removal relies on attention rerouting, which fails on few-step models, and the reason was not known. Matched single-forward interventions on four distilled backbones show that a correction of the clean prediction survives to the final image, whereas attention rerouting acts mainly in later forwards (Tab. 3; App. C). This is specific to the short trajectory: at 50 steps, the same perturbation is attenuated (App. C.4), and ReState transferred unchanged to 50 steps does not work, as you note.
  • Difference from Data Consistency. RePaint, DDNM, and DPS constrain the reconstruction with known observations. In removal, the known observations are the pixels outside the mask, and every row of Tab. 2 already enforces them by re-injecting the source latent there at every step; this alone gives the first row of Tab. 2 (9.86 dB). Inside the mask, the source pixels belong to the object, so data consistency provides no target for the background. The value to be written there must be specified, and this is why ReState has to construct B: it builds a proposal from the unmasked region with the features of the first forward, at no additional denoiser evaluation. Unlike a measurement, B is an estimate, so it is used as soft evidence and not as a constraint; using it as a hard fill is the direct decode of Table R1 (16.55 dB). Sec. 2 now states this difference and positions Eq. (6) as an instance of proximal data-correction whose target is a retrieved pseudo-target, not a measured observation.
  • The Structure of the Schedule Matters. The weight of Eq. (7) is strong while noise remains and zero at the final forward, and both properties are needed. A weak correction (a constant weight of 0.25) loses 1.2 dB (Tab. D2), and keeping the correction at the final forward loses 1.5 dB and raises LPIPS from 0.194 to 0.284 (App. D.2), because the last forward is where fine texture is synthesized. Within this structure, the exact values are not critical: a strong constant weight with the same final release is close to the default, as is the λ sweep of App. D.1. Eq. (7) provides this structure in closed form from the noise level, without a manually tuned schedule.

We hope that the direct-decode experiment, the stronger baselines, and the clarified contribution address your concerns. If any point remains unclear, or if you would like to see further experiments, please let us know; we will respond and add the results as quickly as possible.

AuthorsOpenReview ↗

Additional Updates for Reviewer 1ujD

We thank Reviewer 1ujD again for the careful review. Since our first response, the discussion with the other reviewers has led to several additions to the manuscript. Below, we summarize those that relate to your [W1] and [W3], so that they can be read without the other threads.

# [W1/W3] Method Contribution: A Role for Each Forward of a Few-Step Trajectory

ReState assigns a role to each forward of a few-step trajectory: the first collects background evidence from the input image, those with remaining noise write it into the masked clean prediction, and the final forward keeps object-excluding attention alone. This sentence now states the method contribution in the Introduction. Table R1 compares the default with three variants that keep only part of this design; they are 1.5 to 3.3 dB lower.

  • Relation to [W1]. Correcting the first forward is redundant once the second and third forwards are corrected (18.74 vs. 18.75 dB), as you note. Its role in ReState is to provide the features for retrieval, so that the background evidence is collected at no additional denoiser evaluation.
  • Relation to [W3]. Patch filling alone corresponds to decoding B directly (16.55 dB), and a correction applied at every forward, including the last, gives 17.26 dB with an LPIPS of 0.284. Attention forcing alone (15.50 dB) changes where the masked tokens read from, not what they should become. The default differs from these variants in when the evidence is written and when it is released.

Table R1 (i.e., Tab. 4, Tab. 6, and Tab. D2, Selected Rows). Variants that keep only part of the design (OBER-Test, SDXL-Turbo).

VariantPSNR ↑SSIM ↑LPIPS ↓
B Decoded Directly, No Sampling (Tab. 6)16.550.4480.299
Attention Forcing Only, No Correction (Tab. 4)15.500.3480.269
Correction Kept at the Final Forward (Tab. D2)17.260.4660.284
ReState (Ours)18.750.4980.194

# [W3] Position among Training-Free Removal Methods

  • The Few-Step Setting. AdaEraser reports a clear degradation on four-step SDXL-Lightning and leaves distilled models to future work (their App. D).
  • Source of the Target. Attention-based methods (DesignEdit, Attentive Eraser, AdaEraser, PANDORA) hide the object in self-attention and specify no target for the masked region, and PredErase obtains one from an external model (I-JEPA) through gradient updates. ReState retrieves the target from latents of the same image with the features of the first forward and writes it in closed form, so it needs no second model and no back-propagation, and it adds no denoiser evaluations (Sec. 2; Tab. 1).

# Extension to DiT Backbones (App. G.2)

Beyond the UNet backbones of the submission, the revision applies ReState to five DiT backbones (FLUX.2-klein-4B and 9B, FLUX 3 Image Klein, Qwen-Image-Flash, and Z-Image-Turbo). On OBER-Test, the masked PSNR rises from 10.0–11.3 dB to 17.3–19.6 dB at the native number of forwards of each model (Tab. G3).

We hope that these additions clarify the contribution. If any concern remains, we would be glad to respond within the discussion period.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

Dear Area Chairs and Reviewers,

We sincerely thank the Area Chairs and all reviewers for their time and insightful comments. We are pleased that the reviews acknowledge the practical and well-motivated problem (Reviewer 1ohS), the simple, training-free method that adds no denoiser evaluations (all four reviewers), the carefully controlled intervention analysis that is directly connected to the method (Reviewers CJpo and 1ohS), and the thorough experiments and ablations (Reviewers hdD7, 1ujD, and CJpo).

We revised the manuscript with new experiments and clarifications. Existing section and table numbering in the main paper remains unchanged (with one new table, Tab. 6, added at the end), while most new results are placed in the appendix. Every change is marked in blue. The key updates are:

1. Stronger Baselines

  • [Abstract, Sec. 4, App. E.3, App. B.3; Tab. 5, Tab. E4, Tab. B3, Tab. B4] Official Implementations and Re-Tuned Ports (Reviewers 1ohS, 1ujD, and CJpo): With their official implementations at native step counts, DesignEdit and Attentive Eraser reach 18.51 and 18.19 dB on OBER-Test; 4-step ReState reaches 18.75 dB while being 14.2× and 12.4× faster, and it is also higher than both on RORD and MULAN. Re-tuning the Attentive Eraser port raises it from 7.09 to 15.34 dB.
  • [App. E.1, App. G.1; Tab. E1, Tab. G1] LaMa and Few-Step Editors (Reviewers 1ujD and 1ohS): LaMa is added to Tab. E1. TurboEdit, InfEdit, and ChordEdit stay at 9.29 to 9.95 dB inside the mask, against 18.75 dB for ReState.
  • [Sec. 5, App. D.8; Tab. D7] Combination with SARG and Adaptive Suppression (Reviewer hdD7): Adding the direct correction of ReState raises the re-tuned SARG from 15.34 to 18.57 dB and the adaptive suppression of AdaEraser from 11.71 to 17.79 dB.

2. Evaluation

  • [App. F.1; Tab. F1] Perceptual Metrics, Blur Control, and Figures Without Contours (Reviewer 1ohS): A blur control confirms that masked PSNR rewards smooth fills. On LPIPS, DISTS, CMMD, and ReMOVE, ReState is the best training-free method, and the qualitative figures added in the revision show the outputs without contour overlays.
  • [App. F.2; Tab. F2] Object-Only Benchmark (Reviewer 1ohS): On 500 pairs from MULAN, whose ground truth removes only the object, ReState is 2.6 dB higher than the strongest training-free baseline in masked PSNR and 1.6 dB higher in whole-image PSNR.
  • [App. B, App. F.3; Tab. F3, Tab. F4] Object-Effect Masks and Whole-Image Evaluation (Reviewers 1ohS and CJpo): App. B states the mask used on each benchmark. With the object-effect mask as input, the unchanged ReState improves the effect-only region of OBER-Test by 6.5 dB and the whole image by 1.5 dB.

3. Analysis of the Method

  • [Sec. 4, App. D.3; Tab. 6, Tab. D3] Decoding B Directly (Reviewers 1ujD and 1ohS): Decoding the background evidence B into the mask without denoising gives 16.55 dB with the retrieved B and 15.13 dB with the global mean. Using the same B within ReState gives 18.75 and 18.16 dB.
  • [App. D.4; Tab. D4] One-Shot Insertion of B (Reviewer 1ohS): Inserting B once before the 2nd, 3rd, and 4th forward gives 17.90, 18.53, and 18.61 dB, against 18.75 dB for the per-forward correction. The late insertion is reported as a faster variant.
  • [Sec. 3.4, App. D.2, App. D.7; Tab. D2, Tab. D6] Correction Schedule and Step Count (Reviewers CJpo and 1ujD): A weak correction loses 1.2 dB and a correction kept at the final forward loses 1.5 dB, whereas a strong constant weight with an uncorrected (released) final forward pass performs closely to Eq. (7). With two, three, and four steps, ReState is 3.2 to 3.5 dB above attention forcing alone.

4. Extension and Clarifications

  • [App. A.1, App. G.2; Tab. G2, Tab. G3] DiT Backbones (Reviewer hdD7): Applied to five DiT backbones (FLUX.2-klein-4B and 9B, FLUX 3 Image Klein, Qwen-Image-Flash, and Z-Image-Turbo), ReState raises the masked PSNR on OBER-Test from 10.0–11.3 dB to 17.3–19.6 dB.
  • [Sec. 2, Sec. 3.3; Tab. 1] Related Work (Reviewers 1ohS, CJpo, and 1ujD): We discuss few-step text-driven editing (TurboEdit, InfEdit, and ChordEdit) and PredErase, and state how ReState differs from PANDORA, from classical patch-based filling, and from data-consistency methods.
  • [Sec. 3, Sec. 4, Sec. 5, App. B] Scope and Protocol (Reviewers CJpo, 1ujD, and 1ohS): We state that ReState addresses trajectories with two or more forwards and that the first-forward analysis is a diagnosis of the intervention mechanism, and we fix the broken reference and the captions of Tab. 1 and Tab. 2.

Thank you again for your time; we remain available for any further questions during the discussion.

Sincerely,

Authors of Submission 42004

Dear Reviewers,

The authors have responded to your questions. If you haven’t already done so, please review their responses and let them know whether your concerns have been addressed. Please use the discussion to raise any remaining concerns or follow-up questions.

Thank you for your time!

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.