acceptodds
Under review as a conference paper at ICLR 2027

Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network

Abstract

Text-guided image editing must introduce the requested changes while preserving unrelated source content. In training-free editing, diffusion editors often use spatial controls whose inaccuracies can leave edits incomplete or alter unrelated regions. Causal autoregressive editors face a further constraint: their fixed decoding order limits revision of earlier decisions. As the first to explore training-free image editing with Generative Refinement Networks (GRN), we observe that its refinement process is inherently suitable for editing and offers a promising way to address these limitations. Motivated by this observation, we introduce ***RefineEdit***, a training-free prompt-to-prompt image editing framework built on the GRN. Our key idea is to couple edit localization with content generation through the global refinement of binary image codes, allowing editing evidence to be revised as the image evolves. More specifically, ***RefineEdit*** combines *bit routing* with two stabilization mechanisms: *adaptive spatial freezing* and *finite bit locking*. *Bit routing* starts from an intermediate source state and uses signed probability differences between the two branches to identify editable positions and bits. It directs selected bits toward editing refinement while anchoring the rest to the evolving source trajectory. *Adaptive spatial freezing* limits unnecessary expansion of the editing region, while *finite bit locking* maintains recent bit activations to support continued editing. The overall framework requires no additional training, external masks, or attention control. Across nine editing categories of PIE-Bench, ***RefineEdit*** achieves the best background-preservation scores in PSNR, LPIPS, MSE, and SSIM, together with the highest whole-image and edited-region CLIP scores among the evaluated methods. Code is available at https://anonymous.4open.science/r/RefineEdit.

Primary area:
generative models
Keywords:
Training-Free Image Editing, Text-Guided Image Editing, Generative Refinement Network

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_tts6OpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper introduces RefineEdit, a training-free image editing method based on GRN. It uses bitwise probability differences, source-anchored routing, and two mask stabilization mechanisms to balance editing effectiveness and content preservation. Experiments demonstrate strong background preservation and text alignment.

Critical strengths
  1. The paper presents a clear motivation for leveraging the iterative refinement process of GRN for image editing.
  2. RefineEdit introduces a straightforward editing framework based on bitwise probability differences and source-anchored routing, without requiring additional training, external masks, or attention manipulation.
  3. The method is well-motivated and demonstrates a promising balance between editing effectiveness and content preservation.
Critical weaknesses
  1. My main concern is the results primarily establish that GRN-generated images can be effectively edited through their refinement trajectories, rather than demonstrating general-purpose image editing capabilities. Because both PIE-Bench and EditEval v2 evaluations use GRN-generated source images rather than the original benchmark images, it limits the fairness of the comparisons and its applicability to any real images.
  2. The reported gains may partly stem from differences in the underlying generative backbones rather than the proposed editing strategy.
  3. This work lacks comparisons with recent binary/autoregressive image editing methods, such as AREdit, EditInfinity and BitResEdit, making it difficult to assess its advantages over closely related approaches.
Questions for authors
  1. Both PIE-Bench and EditEval v2 evaluations use GRN-generated source images rather than the original benchmark images. It is very important for the authors to clarify how RefineEdit could be extended to arbitrary real-world images?
  2. Could the authors compare the text-to-image generation capabilities of GRN and the baselines' backbones to clarify whether the reported editing improvements stem from the stronger generative backbone or the proposed editing method?
  3. Could the authors provide comparisons with recent binary/autoregressive image editing methods, such as AREdit, EditInfinity, and BitResEdit, to better establish the advantages of RefineEdit over closely related approaches?
4 comments

Thank you for your thoughtful review and suggestions. We have revised the manuscript and marked the changes in red for ease of content. Our point-by-point responses follow below.

W1 and Q1: Scope of the evaluation and editing real images

Our main task is prompt to prompt editing. a common starting point is an image generated by the base model. RefineEdit reuses an intermediate GRN state from that generation run. Hu et al.[1] used a similar protocol for ISLock. LlamaGen generated the source images from PIE-Bench prompts, and Anchor Token Matching (ATM) used source tokens as anchors during editing. The diffusion and flow baselines receive the same RGB image generated by LlamaGen. In our comparison, every method edits the same GRN image with the same prompt and is scored against the same source. RefineEdit also has access to the source trajectory. Section 5.3 controls for that access.

Paper Appendix C extends RefineEdit to an real image editing given its source and editing prompts. We encode the image once with the frozen HBQ tokenizer and use its bits to construct a source reference at each refinement step. Selected bits follow the editing prompt, and unselected bits copy the reference. This lets us edit an image without its original generation trajectory or an inversion step.

MethodTypeBackground PSNR ↑LPIPS ↓MSE ↓SSIM ↑CLIP Whole ↑CLIP Edited ↑
P2PDiffusion17.870.20880.02200.711425.0122.44
MasaCtrlDiffusion22.170.1066NR0.796723.9621.16
Pix2Pix-ZeroDiffusion20.440.17220.03920.746722.8020.54
PnPDiffusion22.280.11340.00840.790525.4122.55
PnP-DirInvDiffusion22.460.10610.00810.796825.4122.62
ChordEdit (SD-Turbo)Diffusion22.200.12830.007130.759125.5822.96
FlowEditFlow21.960.11220.00950.830825.2522.58
RF-InversionFlow20.820.1900NR0.719225.2022.11
AREditVAR24.190.0870NR0.837025.4222.77
RefineEditGRN25.010.08410.006920.866926.3322.97

We tested this extension on all 700 original PIE-Bench images. RefineEdit achieves the best reported scores on all four background metrics and both CLIP metrics. For the nine named categories, we varied only the switch step and reused the spatial and bitwise thresholds. Tuning these thresholds for real images may further improve the balance between edit strength and preservation.

[1] Hu et al. Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing. ICCV 2025, Sections 3 and 4.1.

W2: What does the editing strategy add to GRN?

GRN was trained to generate images from text. This ability alone does not guarantee that GRN can make a localized edit to image. We ran this experiment to test whether switching prompts midway through refinement is enough to edit a source image while preserving content outside the requested region. In Paper Section 5.3, we compare direct branching, which lets every bit follow the editing prompt, with RefineEdit, which updates selected bits and anchors the rest to the source trajectory. Both use the same GRN and, for each image, the same source state, switch step, and number of remaining steps. We used the Original settings and evaluated on images held out during parameter selection (Appendix B.2).

MethodStructure Distance ↓Background PSNR ↑LPIPS ↓MSE ↓SSIM ↑CLIP Whole ↑CLIP Edited ↑
Direct branching0.0540320.350.11620.014810.783927.0423.76
RefineEdit (full bit routing)0.0266630.360.03180.004570.950426.5923.44

Direct branching follows the editing prompt well (edited CLIP 23.76), but it also changes details outside the target region. Across the evaluation set, routing raises background PSNR by 10.01 dB and lowers background LPIPS from 0.1162 to 0.0318, with a 0.32 decrease in edited CLIP. With the backbone and source state held fixed, this gain in background fidelity shows that selective routing confines edits more reliably than a prompt switch alone. Appendix C shows that the same strategy also works with a source reference constructed from a real image.

W3, Q2 and Q3: Backbone quality and comparisons with autoregressive editors

We measured GRN-2B on GenEval. Its overall score of 0.7036 is close to the published Infinity-2B score of 0.73, which uses prompt rewriting. FLUX.1-dev scores higher than GRN on counting and colors. In Table 1 of our paper, RefineEdit outperforms both FLUX.1-dev editors, FlowEdit and RF-Inversion, on all four background metrics and both CLIP scores.

Bold marks the best score and underline marks the next distinct score in both tables.

GenEval metric↑Infinity-2BGRN-2B (measured)FLUX.1-devSD 1.4SD 1.5
Single object0.98‾\underline{0.98}0.99060.98‾\underline{0.98}0.970.97
Two objects0.850.846‾\underline{0.846}0.810.390.38
Counting0.630.7188‾\underline{0.7188}0.740.330.35
Colors0.830.76860.79‾\underline{0.79}0.730.76
Position0.490.3275‾\underline{0.3275}0.220.030.04
Color attribution0.570.570.45‾\underline{0.45}0.050.06
Overall0.730.7036‾\underline{0.7036}0.660.420.43

We also evaluated AREdit, BitResEdit, and EditInfinity on the GRN source images used for our editing comparison. All three baselines used Infinity-2B in our experiments.

MethodBackboneStructure Distance ↓PSNR ↑LPIPS ↓MSE ↓SSIM ↑CLIP Whole ↑CLIP Edited ↑
AREditInfinity-2B0.0287325.1060.058670.0083820.8778226.42423.057
BitResEditInfinity-2B0.0738826.940.03850.0032320.9322527.26324.377
EditInfinityInfinity-2B0.0750424.7190.041630.0051370.9160926.86724.201
RefineEditGRN-2B0.0266630.360.031780.0045670.9503726.59723.447

Among the four editors, RefineEdit achieves the lowest Structure Distance (0.02666). It also achieves the highest background PSNR (30.36 dB), lowest background LPIPS (0.03178), and highest background SSIM (0.95037). It outperforms AREdit on all seven metrics. Appendix C also compares the two methods on the original PIE-Bench images, where RefineEdit improves PSNR over AREdit by 0.82 dB and SSIM from 0.8370 to 0.8669.

BitResEdit achieves the lowest MSE and the highest CLIP scores. Its external edit mask tells it where to apply the prompt, which may contribute to its higher CLIP scores. RefineEdit doesn't need external edit mask. EditInfinity optimizes a prompt embedding and fits LoRA weights for each input image. RefineEdit requires neither step. Including this optimization, EditInfinity took 2.2 times as long per image as RefineEdit on average in our measurements. Infinity-2B slightly outperforms GRN-2B on GenEval overall (0.73 vs. 0.7036). Despite this gap, RefineEdit outperforms all three Infinity-2B editors on Structure Distance, PSNR, LPIPS, and SSIM, demonstrating the effectiveness of its editing strategy in preserving scene structure and background detail.

I appreciate the authors’ response. The real-image editing demonstrations have eased my doubts about practical usability. I recommend adding comparisons and analyses with Infinity-based editing methods, given the relevant connections between the proposed method and this line of work. Specifically, I would prefer to see editing comparisons on the real-image of PIE-Bench instead of GRN-generated images. Furthermore, I believe comparisons with several state-of-the-art image editing models, such as FLUX.1 Kontext and Qwen-Image-Edit, could enrich the paper. Nevertheless, I view this editing method built upon the novel GRN architecture as a substantial contribution to the community, and it also exhibits promising performance on real-image editing tasks; I will therefore increase my score to 4.

Thank you very much for your positive update and thoughtful suggestions. We are pleased that our real-image editing demonstrations helped address your concerns. We hope our work can offer interesting insights to the community and inspire further explorations on training-free image editing models.

Reviewer_9ajsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RefineEdit, a training-free image editing framework based on a Generative Refinement Network. It employs a globally modifiable binary representation to simultaneously localize the regions for editing and refine their content. The method identifies target editing areas by calculating the signed probability difference between a source branch and an editing branch. It then utilizes a bit-routing mechanism to perform localized edits while preserving unaffected content. Furthermore, the framework incorporates Adaptive Spatial Freezing and Finite-Bit Locking mechanisms to suppress the expansion of edited regions and mitigate probability fluctuations during the editing process. Experiments on PIE-Bench and EditEval v2 demonstrate that the proposed method excels in both background preservation and alignment with text-based semantic intent.

Critical strengths
  1. The primary strength of this paper lies in its effective utilization of the globally modifiable binary representation (GRN) to propose a training-free image editing method with a clear design rationale. Through a dynamic bit-routing mechanism based on probabilistic discrepancies, RefineEdit adaptively adjusts editing decisions during the iterative generation process, while preserving original content by anchoring unselected bits to a continuously updated source-image trajectory. Furthermore, AdaSF and FBL optimize the editing process by addressing spatial constraints and temporal continuity, respectively. Tailored specifically to the unique generative mechanism of GRN, this approach demonstrates both innovation and sound reasoning.
  2. Experimental results demonstrate that RefineEdit is highly competitive on the PIE-Bench benchmark, exhibiting a strong ability to preserve the background. Furthermore, the method achieves the highest CLIP similarity scores—both for the entire image and within the edited regions—indicating an effective balance between the desired edits and the retention of original content. The authors also validate the method’s efficacy through cross-dataset experiments using EditEval v2, as well as through ablation studies and user preference evaluations.
Critical weaknesses
  1. Equation (3) uses probability differences between the source and editing branches to identify regions for editing. However, after the switch step, the branches also have different latent states, and the paper does not isolate how these state differences affect edit localization. The choice of score also needs further examination. Since it relies on source-sampled bits, its expected value is E[Δ]=(2ps−1)(ps−pe)\mathbb{E}[\Delta] = (2p_s - 1)(p_s - p_e), which can vanish even when the predictive distributions differ substantially. Averaging signed differences across channels in Equation (4) may also cancel out meaningful changes. Comparisons with alternative scoring functions and a shared-state control would help assess these choices.
  2. Both PIE-Bench and EditEval v2 use GRN-generated source images, and RefineEdit requires access to their generation trajectories. Although the diffusion and flow baselines edit the same images, they do not have access to these intermediate states. The reported results therefore establish the method's effectiveness in the GRN setting, but its applicability to real images and its broader advantage over other editing methods remain unclear.
Questions for authors
  1. How does RefineEdit compare with simply switching to the editing prompt at the same GRN refinement step, without bit routing or mask stabilization?
  2. Have you tried computing the probability differences in Equation (3) under a shared latent state? It would be useful to see whether the divergence between the two branches helps localize edits or introduces unwanted changes.
  3. Why use signed probability differences in Equation (3) rather than absolute differences or other distributional measures? Since positive and negative values can cancel when averaged across channels in Equation (4), how does this affect spatial mask selection?
  4. AdaSF freezes the spatial mask when the initial editing response is strong, but this does not guarantee that the entire target region has been identified. Have you observed cases where early freezing leads to incomplete edits?
  5. Were the category-specific hyperparameters tuned on a separate validation split or on the same PIE-Bench samples used for the final evaluation?
3 comments

Thank you for your thoughtful review and suggestions. We have revised the manuscript and marked the changes in red. We respond to each point below.

W1 and Q3. Why use signed probability differences?

Let psp_s and pep_e denote the probabilities of a one in the source and editing branches. Then E[Δ∣ps,pe]=(2ps−1)(ps−pe)\mathbb{E}[\Delta\mid p_s,p_e]=(2p_s-1)(p_s-p_e). In our rerun, we examined about 2.621 billion bit observations across refinement steps for which ∣ps−pe∣|p_s-p_e| exceeded the category's bit threshold. Only 95 had a conditional expectation of exactly zero at the recorded precision. Values close to zero do occur when the source bit is uncertain. An exact zero at the recorded precision was extremely rare, occurring in about one in 28 million of the observations examined. Exact zeros therefore account for a negligible fraction of the observations above the bit threshold. We therefore also checked how often spatial positions passed selection.

We reapplied both rules to the same scores along the original Signed trajectory. The table gives the fraction of spatial positions that passed the threshold.

Position relative to the reference edit maskSignedAbsolute
Inside57.94%76.79%
Outside18.49%48.73%

Signed admitted fewer positions inside the reference region, as the reviewer anticipated. Allowing edits at more of these positions with Absolute improved the edited-region CLIP score by only 0.011. Absolute also admitted far more positions outside the target (48.73% versus 18.49% with Signed, a difference of 30.24 percentage points). This came at a substantial cost to background fidelity in the final images. Background PSNR fell from 30.36 dB with Signed to 23.83 dB with Absolute, a drop of 6.53 dB.

We also ran each rule through the full editing process. The last column gives the reported difference (Absolute minus Signed). Rounding can affect the last digit.

MetricSigned (RefineEdit)AbsoluteDifference
Structure Distance↓0.026660.03895+0.01229
Background PSNR↑30.3623.83−6.53 dB
Background LPIPS↓0.031780.06489+0.03311
Background MSE↓0.004570.00852+0.00396
Background SSIM↑0.950370.87649−0.07389
CLIP Whole↑26.59726.615+0.018
CLIP Edited↑23.44723.458+0.011

In full runs, opening more positions gained just 0.011 in edited CLIP, while background PSNR fell by 6.53 dB and LPIPS more than doubled. This is why we retain the sign despite the lower selection rate inside the target region.

W2. Generated sources and applicability to real images

Our main task is prompt to prompt editing. A user generates a source image and then revises it with a new prompt. Hu et al.[1] used a similar protocol for Anchor Token Matching (ATM), generating source images with LlamaGen from PIE-Bench prompts before editing them with ISLock. The diffusion and flow baselines receive the same RGB image generated by LlamaGen. In our main comparison, every method edits the same GRN image with the same prompt. RefineEdit also has access to the GRN source trajectory.

During rebuttal, we extended RefineEdit to edit real images directly and evaluated this extension on all 700 original PIE-Bench images (Appendix C). We encode the input once with the frozen HBQ tokenizer and combine its binary code with a fixed random code along the GRN refinement schedule. This gives the editor a source reference without inversion or access to the original generation trajectory. At the final step, selected bits follow the editing proposal and all other bits copy the input code. The table below reproduces its quantitative results. Baseline scores are previously reported values. NR means the value was unreported.

MethodTypeBackground PSNR↑LPIPS↓MSE↓SSIM↑CLIP Whole↑CLIP Edited↑
P2PDiffusion17.870.20880.02200.711425.0122.44
MasaCtrlDiffusion22.170.1066NR0.796723.9621.16
Pix2Pix-ZeroDiffusion20.440.17220.03920.746722.8020.54
PnPDiffusion22.280.11340.00840.790525.4122.55
PnP-DirInvDiffusion22.460.10610.00810.796825.4122.62
ChordEdit (SD-Turbo)Diffusion22.200.12830.007130.759125.5822.96
FlowEditFlow21.960.11220.00950.830825.2522.58
RF-InversionFlow20.820.1900NR0.719225.2022.11
AREditVAR24.190.0870NR0.837025.4222.77
RefineEditGRN25.010.08410.006920.866926.3322.97

RefineEdit achieved the best reported score on all four background metrics and both CLIP metrics. Its edited CLIP was close to ChordEdit's (22.97 vs. 22.96), while its background PSNR was 2.81 dB higher.

[1] Hu et al. Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing. ICCV 2025, Sections 3 and 4.1.

Q1. What happens if GRN simply switches prompts?

Direct branching starts from the source state at the switch step and lets every bit follow the editing prompt, without routing or mask stabilization. Both runs use the same GRN, source trajectory, switch step, and number of remaining refinement steps. The Original settings and details can be found in Section 5.3 and Appendix B.2.

MethodStructure Distance↓Background PSNR↑LPIPS↓MSE↓SSIM↑CLIP Whole↑CLIP Edited↑
Direct branching0.0540320.350.11620.014810.783927.0423.76
RefineEdit0.0266630.360.03180.004570.950426.5923.44

Direct branching follows the new prompt, but the change extends beyond the target. Across the evaluation set, routing improves background PSNR by 10.01 dB, with a 0.32 decrease in edited CLIP.

Q2. Scoring both prompts on a shared state.

We tested this control on all PIEBench images. In Shared-state Signed, both prompts are scored on the current source state in Eq. (3), while the editing branch still proposes updates from its own state. In Original Signed, each prompt is scored on its own branch state. Both runs used the same settings.

The table shows the percentage of bits copied from the source branch at each refinement step. We measured exact matches to the whole source image at the final step.

Refinement stepOriginal SignedShared-state Signed
t=20t=2090.94%98.01%
t=30t=3086.83%99.34%
t=40t=4083.17%99.75%

As the source image took shape, changing the prompt had less effect on the predicted bits. By step 40, only 0.25% of bits could still follow the editing proposal. Shared-state Signed reproduced the source image exactly in 43.57% cases (no editing), compared with 0.54% for Original Signed.

As refinement progresses, the editing prompt has less influence when both prompts are scored on the same source state. The editable region narrows, and the editing branch copies more source bits. When all bits come from the source, a request to turn Spider-Man into a robot would simply reproduce the original Spider-Man image. Scoring the editing prompt on its own evolving state helps RefineEdit carry the requested change through the final steps.

Q4 Early freezing

Yes. A strong response in one part of an object sometimes froze the mask before it covered the rest. A direction for future work is to check whether the mask is still growing before freezing it.

Q5. Which images were used for parameter selection?

For the Original settings, we randomly selected five images from each of the nine PIE-Bench categories, giving 45 validation images. We ran a small parameter search on these images only. Table 1 reports results on the remaining images.

During rebuttal, we drew five new images per category, excluding all 45 from the first set, and repeated selection on these 45 images. We evaluated the Reselected settings, the Original settings, and the external baselines on the same images outside the second set. Six of the nine categories retained exactly the same parameter values. The reselection also used a different selection rule. Appendix B.2 describes both rules and the search ranges in red.

Parameter comparison. Each triple lists the switch step, spatial threshold, and bitwise threshold, in that order.

Editing categoryOriginalReselected
Object replacement(15,0.013,0.20)(15,0.013,0.20)
Object addition(15,0.016,0.10)(15,0.016,0.10)
Object removal(10,0.020,0.16)(10,0.020,0.16)
Content modification(12,0.016,0.18)(12,0.016,0.18)
Pose modification(12,0.018,0.16)(12,0.018,0.16)
Color modification(11,0.020,0.16)(13,0.018,0.10)
Material modification(16,0.020,0.20)(16,0.020,0.12)
Background modification(10,0.012,0.16)(10,0.012,0.16)
Style transfer(6,0.016,0.14)(10,0.010,0.10)

Editing results on the images outside the second selection set. Bold and underline mark the best and second-best of the three scores shown for each metric.

MetricStrongest external baselineBaseline scoreRefineEdit (Original)RefineEdit (Reselected)
Structure Distance↓FlowEdit0.016220.026660.02245‾\underline{0.02245}
PSNR↑FlowEdit25.0330.3630.25‾\underline{30.25}
LPIPS↓FlowEdit0.061680.031780.03196‾\underline{0.03196}
MSE↓FlowEdit0.0045980.0045670.004575‾\underline{0.004575}
SSIM↑FlowEdit0.901960.950370.94988‾\underline{0.94988}
CLIP Whole↑LEDits++26.21526.59226.468‾\underline{26.468}
CLIP Edited↑LEDits++23.28923.44223.302‾\underline{23.302}

With Reselected, RefineEdit still outperformed every external baseline on all four background metrics and both CLIP metrics. Structure Distance improved from 0.02666 to 0.02245.

The Original settings also transferred directly to the 150 EditEval v2 prompt pairs without retuning (Appendix D.4).

Thank you again for your time and thoughtful feedback. With two days remaining in the discussion period, please let us know if any questions or concerns remain after our responses. We would be happy to provide further clarification.

Reviewer_G84MOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RefineEdit branches from an intermediate GRN source-generation state and uses source/editing probability differences for source-sampled bits to select spatial and bitwise editing permissions. Unselected coordinates follow the source trajectory, while adaptive spatial freezing and finite bit locking stabilize permissions. Evaluation uses 560 PIE-Bench and 150 EditEval v2 prompt pairs with newly generated GRN sources, rather than the original benchmark photographs. The paper reports strong preservation and competitive alignment, mechanism ablations, and a 20-example human study. EditEval v2 reuses category settings without retuning.

Critical strengths
  1. An interesting editing interface for a revisable generative representation. Aligned source/editing bit coordinates allow localized changes without extra training, and combining refinement with editing permissions is a useful direction beyond fixed-order decoding.

  2. Supporting evidence extends beyond the primary prompt set. The additional 150-prompt evaluation transfers category settings without retuning, and the human study provides preference evidence on 20 example sets (Appendices C and D.4). These results reduce concern that the method only works on selected qualitative examples.

Critical weaknesses
  1. The main evaluation is involved in parameter selection. Appendix A.3 states that category-specific hyperparameters are selected using PIE-Bench evaluation results, and Table 4 gives separate switch/spatial/bit settings for nine categories. Without an independent selection split, the primary comparison is not a clean held-out estimate. Untuned EditEval v2 results are useful mitigation, but do not remove selection bias from the main results. This materially limits my confidence in the reported comparative gains.

  2. The editing contribution is not isolated from native trajectory access and the backbone. RefineEdit has GRN intermediate states, whereas other editors receive the same final RGB image. Sharing the source does not equalize internal information. Direct branching, spatial-only routing, and full bit routing are not compared within the same GRN setting. The stabilization ablations show a preservation–editing trade-off, but do not establish a better trade-off than simpler same-backbone routing. This makes the incremental mechanism contribution uncertain; lack of arbitrary-real-image support is not itself my reason for rejection.

Questions for authors
  1. Clarification of evaluation independence: Which examples, objectives, and search budget were used for category-specific selection? Was any subset held out from selection? If not, can you freeze settings selected on a validation subset and report the existing task on an independent subset?

  2. Bounded validation of the mechanism claim: Can you compare direct branching, spatial-only routing, and full bit routing using the same GRN source trajectories and budget, with settings selected independently of evaluation? Such a comparison could change my recommendation. Extending the method to arbitrary photographs is future work, not a required rebuttal experiment.

5 comments

Thank you for your thoughtful review and suggestions. We have revised the manuscript and marked the changes in red for ease of content. Our point-by-point responses follow below.

W1 and Q1: Which images were used for parameter selection?

For Original paper, we randomly drew five images from each of the nine PIE-Bench categories (45 in total). We chose the parameters for each category using only these images. Paper Table 1 reports results on the remaining images.

During rebuttal, we drew another five images per category, excluding all 45 from the first set. We used this second set to choose the Reselected settings and compared both configurations and all baselines on the images outside this set. In both runs, we searched a narrow range of parameter values on the 45 validation images. In paper, Appendix B.2 gives the selection rules and search ranges in red.

Parameter comparison. Six of the nine categories kept exactly the same settings. Each triple lists the switch step, spatial threshold, and bitwise threshold, in that order.

Editing categoryOriginalReselected
Object replacement(15, 0.013, 0.20)(15, 0.013, 0.20)
Object addition(15, 0.016, 0.10)(15, 0.016, 0.10)
Object removal(10, 0.020, 0.16)(10, 0.020, 0.16)
Content modification(12, 0.016, 0.18)(12, 0.016, 0.18)
Pose modification(12, 0.018, 0.16)(12, 0.018, 0.16)
Color modification(11, 0.020, 0.16)(13, 0.018, 0.10)
Material modification(16, 0.020, 0.20)(16, 0.020, 0.12)
Background modification(10, 0.012, 0.16)(10, 0.012, 0.16)
Style transfer(6, 0.016, 0.14)(10, 0.010, 0.10)

Editing results. All scores below use the images outside the second 45-image selection set. The baseline column gives the strongest external method for each metric. Bold and underline indicate the best and second-best of the three scores shown.

MetricStrongest external baselineBaseline scoreRefineEdit (Original)RefineEdit (Reselected)
Structure Distance↓FlowEdit0.016220.026660.02245‾\underline{0.02245}
PSNR↑FlowEdit25.0330.3630.25‾\underline{30.25}
LPIPS↓FlowEdit0.061680.031780.03196‾\underline{0.03196}
MSE↓FlowEdit0.0045980.0045670.004575‾\underline{0.004575}
SSIM↑FlowEdit0.901960.950370.94988‾\underline{0.94988}
CLIP Whole↑LEDits++26.21526.59226.468‾\underline{26.468}
CLIP Edited↑LEDits++23.28923.44223.302‾\underline{23.302}

With Reselected, RefineEdit still outperformed every external baseline on all four background metrics and both CLIP metrics. Structure Distance improved from 0.02666 to 0.02245, with FlowEdit still best at 0.01622.

The submitted paper also reports direct transfer of the Original settings from PIE-Bench to EditEval v2 without retuning (Appendix D.4).

W2 and Q2: Direct branching / Spatial-only / Full bit & Real Image Editing

Why use generated source images? In prompt to prompt editing, a common starting point is an image generated by the base model. RefineEdit reuses an intermediate GRN state from that generation run. Hu et al.[1] used a similar protocol for ISLock. LlamaGen generated the source images from PIE-Bench prompts, and Anchor Token Matching (ATM) used source tokens as anchors during editing. The diffusion and flow baselines receive the same RGB image generated by LlamaGen. In our comparison, every method edits the same GRN image with the same prompt and is scored against the same source.

What does routing add? Direct branching switches to the editing prompt and lets every bit follow the editing proposal. Spatial-only lets every bit at a selected token follow the edit. RefineEdit selects individual bits there and copies the source elsewhere.

For each image, the three variants share the GRN source trajectory, switch step, and number of remaining refinement steps. We used the same Original category settings for all applicable parameters, with no further tuning. We evaluated on the images outside the 45 used for parameter selection (Section 5.3).

MethodStructure Distance ↓Background PSNR ↑LPIPS ↓MSE ↓SSIM ↑CLIP Whole ↑CLIP Edited ↑
Direct branching0.0540320.350.11620.014810.783927.0423.76
Spatial-only0.0311427.460.04490.006180.923926.5723.40
RefineEdit (full bit routing)0.0266630.360.03180.004570.950426.5923.44

Direct branching gives the highest CLIP scores, yet its background PSNR is 10.01 dB below RefineEdit. The intermediate state lets GRN generate content aligned with the editing prompt, but changes extend into the background. Spatial-only improves background preservation. Selecting individual bits adds another 2.90 dB in PSNR while edited CLIP stays close (23.44 vs. 23.40). Figure 4 in Section 5.3 of the revised manuscript shows the qualitative results. Compare the blue and red boxes with the source in column (a). Full RefineEdit preserves these background details most faithfully.

Can RefineEdit edit an existing real image? During rebuttal, we extended RefineEdit to edit real images directly and evaluated this extension on all 700 original PIE-Bench images (Appendix C). We encode the input once with the frozen HBQ tokenizer and combine its binary code with a fixed random code along the GRN refinement schedule. This gives the editor a source reference without inversion or access to the original generation trajectory. At the final step, selected bits follow the editing proposal and all other bits copy the input code.

The table reproduces Appendix C. Baseline values are previously reported results. For RefineEdit, the two CLIP scores use all 700 images, and the background metrics use the 554 images with valid scores. Bold marks the best reported value. NR means the value was not reported.

MethodTypeBackground PSNR ↑LPIPS ↓MSE ↓SSIM ↑CLIP Whole ↑CLIP Edited ↑
P2PDiffusion17.870.20880.02200.711425.0122.44
MasaCtrlDiffusion22.170.1066NR0.796723.9621.16
Pix2Pix-ZeroDiffusion20.440.17220.03920.746722.8020.54
PnPDiffusion22.280.11340.00840.790525.4122.55
PnP-DirInvDiffusion22.460.10610.00810.796825.4122.62
ChordEdit (SD-Turbo)Diffusion22.200.12830.007130.759125.5822.96
FlowEditFlow21.960.11220.00950.830825.2522.58
RF-InversionFlow20.820.1900NR0.719225.2022.11
AREditVAR24.190.0870NR0.837025.4222.77
RefineEditGRN25.010.08410.006920.866926.3322.97

RefineEdit leads on all six metrics, with similar edited CLIP to ChordEdit (22.97 vs. 22.96) and 2.81 dB higher background PSNR.

[1] Hu et al. Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing. ICCV 2025, Sections 3 and 4.1.

Thank you again for your time and thoughtful feedback. With two days remaining in the discussion period, please let us know if any questions or concerns remain after our responses. We would be happy to provide further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.