acceptodds
Under review as a conference paper at ICLR 2027

FreEdit: One-Step Source Memory for Mask-Guided Image Editing

Abstract

Text-to-image (T2I) models can also edit real images according to text prompts. Training-free methods often rely on inversion to preserve source content, requiring repeated evaluations at multiple noise levels and full-image denoising during generation. This is particularly costly for local editing, where the edit region is known in advance and the remaining content should stay unchanged. We present FreEdit (Free-inversion Editing), a training-free and highly efficient method that eliminates multi-step inversion and global denoising for local image editing. FreEdit adds noise to the encoded source at a single intermediate timestep and extracts its attention projections with only one network evaluation. During target denoising, only masked image tokens are updated, attending to stored source keys and values outside the mask for contextual guidance. The generated foreground is combined with the clean source latent, preserving untouched content by construction. For color and style edits, optional QK-CFG further improves source-structure preservation while keeping foreground appearance target-driven. Dedicated appearance evaluations demonstrate strong color fidelity, structural preservation, style alignment, and aesthetic quality. On FLUX.1-dev, FreEdit achieves a 3.58× speedup on the same hardware while improving all six reported mean metrics. Results across Klein, Qwen, and SD3.5 further demonstrate its consistent effectiveness across diverse generative backbones.

Primary area:
generative models
Keywords:
Training-free, Image editing, Text-to-Image

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nySdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method called FreEdit, a training-free method for local, mask-guided image editing on rectified-flow DiTs. Instead of multi-step inversion, it adds noise to the source latent at one time-step, runs a single forward pass, and caches the source attention Q/K/V. During target denoising, only masked tokens are updated. They attend to the cached source K/V outside the mask and to the current target K/V inside it. The result is then composited into the clean source latent. The paper reports results on PieBench benchmarks comparing with previous methods.

Critical strengths

None

Critical weaknesses
  • Limited technical novelty. Injecting source-branch attention features into the target branch is not new and it has been well established in editing, inpainting, and personalization, e.g., P2P, Plug-and-Play, MasaCtrl, KV-Edit, and ReFlex. However, FreEdit does not differentiate their core design of QKV injection with KV-Edit and ReFlex.
  • The “one-step” framing is misleading. Only the source read is one step. The target branch still runs multi-denoising steps, so the total cost is far from one- or few-step editing. The title and abstract invite readers to read this as a one-step editor. The paper should report total NFEs and position the method against other previous few-step editors.
  • Missing citations and comparisons with one/few-step editing and inpainting. Prior one-step and few-step editing and inpainting methods, such as ChordEdit, TurboEdit, SwiftEdit, and InverFill, are neither cited nor compared. These are the essential efficiency baselines, given that the paper’s main claim is speed.

References:

MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing

Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation

ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation

ChordEdit: One-Step Low-Energy Transport for Image Editing

SwiftEdit: Lightning Fast Text-guided Image Editing via One-step Diffusion

InverFill: One-Step Inversion for Enhanced Few-Step Diffusion Inpainting

TurboEdit: Instant text-based image editing

Questions for authors

Please differentiate the core technical contribution compared to previous works on using attention hijacking for image editing

4 comments

We thank the reviewer for the careful and constructive review.

1. Technical distinction from attention-based editing (weakness 1 and question)

We agree that source-feature injection and attention control are established ideas, including in P2P, Plug-and-Play, MasaCtrl, KV-Edit, and ReFlex. We do not claim attention injection itself as our contribution. In particular, foreground-only queries, background K/V reuse, and source-background latent composition overlap with KV-Edit and should be acknowledged explicitly.

Our contribution concerns constructing the source memory with one model evaluation and reusing it throughout a mask-restricted target trajectory. FreEdit reads source features once at a fixed intermediate noise level; it does not construct an inversion trajectory or recompute a time-dependent source branch at each editing step. The stored layer-wise memory supplies background attention while target computation updates editable tokens.

We also agree that ReFlex is an important comparison. ReFlex extracts intermediate source features, so intermediate feature extraction alone is not a unique contribution of FreEdit. The relevant differences include the computation used to reach and process the source state, the subsequent source evaluations, and how memory is reused during foreground editing. We have now evaluated the official ReFlex FLUX implementation, adapted to the common 512 × 512 resolution and supplied masks while retaining its default algorithm settings, under our existing 620-example protocol:

MethodHPS ×100 ↑BG PSNR ↑BG LPIPS ×1000 ↓BG MSE ×10000 ↓Edit CLIP ×100 ↑IR ×10 ↑
ReFlex28.129031.031030.911813.545623.22988.8237
FreEdit (paper configuration)28.569133.822212.47816.778823.71969.5364

FreEdit obtains better values on all six metrics in this evaluation. We will add this comparison and expand the technical discussion, separating shared attention mechanisms from the source-computation and memory-reuse design.

2. Meaning of “one-step” and total computation (weakness 2)

The title is “FreEdit: One-Step Source Memory for Mask-Guided Image Editing”: “one-step” qualifies source memory, not the complete editing process. The target branch still performs multi-step denoising. We agree that this distinction must also be explicit in the abstract and efficiency discussion, and we will clarify it rather than describe FreEdit as a one-step image editor.

For a closer comparison with inversion-free KV-Edit, we additionally evaluate a FreEdit configuration with more closely aligned mask handling and latent composition, while retaining our single-source-read mechanism. We denote this as FreEdit (comparison configuration). The ReFlex quality comparison above uses the paper configuration.

We audited the executed schedules to report source and target evaluations separately:

MethodSource sample evaluationsTarget sample evaluationsTotal sample evaluationsTransformer calls
FreEdit (paper / comparison configuration)1272828
KV-Edit, inversion-free28285656
ReFlex49287749

Here, sample evaluations count each source or target member of a batch separately. ReFlex has 21 source-only evaluations followed by 28 batched source/target calls, explaining the difference between its sample evaluations and Transformer calls. These counts exclude text encoding and VAE operations. Since token counts and batching differ, NFE is not a substitute for measured runtime.

For the additional 620-example efficiency comparison, FreEdit (comparison configuration) takes 3.8406 s/image versus 9.2604 s/image for inversion-free KV-Edit, on RTX 4090 48 GB hardware. This measures the benefit of reducing source processing within multi-step editing.

3. Relationship to one-/few-step editing and inpainting (weakness 3)

We appreciate these references and agree that they belong in the broader efficiency discussion: ChordEdit, TurboEdit: Instant text-based image editing, SwiftEdit, and InverFill. We will discuss them and distinguish the source-processing budget from the target-denoising budget.

One-/few-step editors address the broader goal of low-latency editing, so they are relevant context. However, the specific question studied here is whether source processing and memory construction can be reduced while retaining a multi-step FLUX.1-dev editor. Comparing against inversion-free KV-Edit and ReFlex on the existing backbone and evaluation protocol addresses that question more directly. A comparison using a different generative model or target-step budget would also measure those differences, rather than isolate the source-memory construction.

InverFill also reduces source-processing cost, but does so by training an inversion network to predict a noise initialization for few-step inpainting with SDXL-Turbo or SANA-Sprint. FreEdit instead requires no additional training: a single evaluation of the pretrained FLUX.1-dev model constructs attention memory that is reused throughout multi-step target denoising. Thus, the two methods differ in both what the single source operation produces—an initialization versus reusable attention memory—and the training and sampling setup. InverFill is relevant as a system-level efficiency reference, whereas KV-Edit and ReFlex provide more direct comparisons of source processing within the same FLUX.1-dev backbone.

We therefore focus the additional experiments on KV-Edit and ReFlex to evaluate source-memory construction with the same FLUX.1-dev backbone. We will discuss the cited one-/few-step methods as complementary approaches to efficient editing and clarify that our efficiency comparison concerns the evaluated multi-step setting.

Reviewer_nySdOpenReview ↗

Official Comment by Reviewer nySd

I thank the authors for their rebuttal. The response partially addresses my concerns, particularly by clarifying the one-step framing and adding the missing discussion of recent work.

Although the novelty remains limited, I believe this work might be valuable to the image editing community. Since FreEdit claims efficiency as a key contribution, I encourage the authors to include further comparison and discussion with recent advances in efficient image editing and inpainting, such as ChordEdit, TurboEdit, SwiftEdit, InverFill, and TurboFill, to better contextualize FreEdit in the final version.

I am raising my score to Weak Accept.

Thank you for reconsidering our work and for your positive recommendation. We appreciate your suggestions for strengthening its positioning. The revised manuscript now includes the additional experimental results and a discussion of ChordEdit, TurboEdit, SwiftEdit, InverFill, and TurboFill in Appendix J. We will retain these additions and further refine the comparison and discussion in the final version, clarifying their relationship to FreEdit in source processing, training requirements, and target generation.

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

The paper proposes a training-free method for explicit mask-guided image editing. One evaluation of the source at an intermediate noise level provides fixed attention memory for foreground-only target denoising. An optional QK-CFG rule supports appearance edits, and final latent composition retains the source background. Validation includes a 620-image PIE-Bench subset, controlled source-computation ablations, four-seed stability, appearance evaluations, and transfer to additional backbones. A matched ablation reduces source evaluations from 27 to one and runtime from 15.22 to 3.89 seconds with comparable overall quality.

Critical strengths
  1. A consequential efficiency result with a direct control. Table 4(b) and Appendix C.4 hold the target editor, initialization, masks, and final composition fixed while replacing full source inversion with one source evaluation. The large runtime reduction with comparable overall quality isolates the central contribution and is my main reason to support acceptance.

  2. Supporting evidence beyond a single configuration. Extraction-time sensitivity, four-seed stability, a source-prompt ablation, and transfer to Klein, Qwen, and SD3.5 strengthen the case that reusable source memory is a practical finding rather than a fragile example (Appendices B–D).

Critical weaknesses
  1. The practical advance over the closest prior variant is not fully established. KV-Edit already uses foreground-only queries, background K/V, and source-background latent composition, and includes an inversion-free variant. The distinctive contribution is therefore primarily the single, fixed source read. The main comparison does not separately evaluate inversion-free KV-Edit under the same protocol. This matters to my assessment of the contribution's size: matching an existing variant in both quality and cost would weaken the case, even though the controlled full-inversion ablation remains useful.
Questions for authors
  1. Clarification: Which KV-Edit variant and settings are used in the reported quality and runtime comparisons? How does your single-source-read construction differ from its inversion-free implementation in source-model evaluations and memory reuse?

  2. Bounded validation of the current claim: Can you compare the inversion-free variant on the existing evaluation protocol and hardware, reporting quality, source/target evaluations, and runtime? This comparison could materially change my assessment of the practical contribution. I am not requesting a new application or benchmark.

3 comments

We thank the reviewer for the careful and constructive review.

1. KV-Edit variant and the source-memory distinction (question 1)

We agree that inversion-free KV-Edit is the closest additional comparison and should have been evaluated explicitly. The original main-table KV-Edit quality row is the published inversion-based comparison, not the inversion-free result newly reported below. The additional result below is labeled KV-Edit, inversion-free.

The distinction is not foreground-only queries, background K/V reuse, or latent composition individually: these are shared with KV-Edit. It is the source computation needed to supply that memory. FreEdit evaluates the source model once at a fixed intermediate noise level and reuses the resulting layer-wise memory throughout target denoising. In the evaluated inversion-free KV-Edit route, source features are recomputed at successive editing timesteps. Avoiding inversion therefore does not imply using a single source-model evaluation.

The inversion-free KV-Edit schedule uses 28 source and 28 target evaluations, whereas FreEdit uses one source evaluation and 27 target evaluations.

2. Quality and cost against inversion-free KV-Edit (weakness and question 2)

We completed the comparison on the paper’s 620-image PIE-Bench subset, using the same six quality metrics. For a closer comparison with inversion-free KV-Edit, we additionally evaluate a FreEdit configuration with more closely aligned mask handling and latent composition, while retaining our single-source-read mechanism. We denote this as FreEdit (comparison configuration).

MethodHPS ×100 ↑BG PSNR ↑BG LPIPS ×1000 ↓BG MSE ×10000 ↓Edit CLIP ×100 ↑IR ×10 ↑
KV-Edit, inversion-free28.069835.18039.41605.087223.01617.6896
FreEdit (comparison configuration)28.411234.926610.61135.620723.46769.1160

All methods follow the same PIE-Bench evaluation protocol.

Compared with inversion-free KV-Edit, FreEdit achieves higher HPS, Edit CLIP, and ImageReward with comparable background preservation.

We also collected per-image timing over all 620 examples on RTX 4090 48 GB hardware:

MethodSource NFETarget NFETotal NFETime/image ↓Peak memory ↓
KV-Edit, inversion-free2828569.2604 s34.0898 GiB
FreEdit (comparison configuration)127283.8406 s34.2988 GiB

We follow the paper’s timing protocol: runtime is averaged over all 620 examples, from input to saved output, excluding model initialization. NFE counts source and target model evaluations.

FreEdit (comparison configuration) is 2.41× faster, with comparable peak memory. It combines comparable background preservation with higher HPS and text-alignment scores at substantially lower computational cost. We will include this comparison and clarify the source-memory distinction.

Thank you for your rebuttal and for taking the time to address my concerns. Your responses have answered most of my questions and provided useful clarifications. After considering the rebuttal, I will maintain my score of 3.

Thank you for your follow-up and continued support. We appreciate the time you have taken to evaluate our work and discuss the additional results, and we are glad that our responses have answered your questions and provided useful clarifications.

Reviewer_6w7QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FreEdit, a training-free method that efficiently performs local image editing by eliminating the costly multi-step inversion and global denoising required by existing text-to-image models. It achieves efficient, localised editing by starting with a single network evaluation and updating only masked image tokens during target denoising. The stored source latent is used to ensure that untouched content is preserved.

Critical strengths

The one-step read and QK-CFG are novel and concrete contributions.

Critical weaknesses
  1. Some of the core contributions of this paper are close to those of KV-Edit. However, the latter is only briefly introduced in the related work section as "reuses source-background key-value features". Similar contributions include: compute only editable/foreground tokens; reuse source/background KV cache based on a mask; use only target/foreground queries in attention; combine the generated foreground with the original background for preservation. The paper should discuss more about the relationship between the proposed method and KV-Edit.

  2. The effectiveness of QK-CFG is not clearly evaluated. QK-CFG is disabled for the quantitative semantic evaluation. The colour and style configurations use fixed QK-CFG schedules.

Questions for authors
  1. Please compare the proposed method with KV-Edit thoroughly.

  2. The QK-CFG sounds novel, but its effectiveness needs to be isolated in an ablation study.

2 comments

We thank the reviewer for the careful and constructive review.

1. Relationship to KV-Edit (weakness and question 1)

We agree that our original discussion of KV-Edit was too brief. Foreground-only queries, mask-based background K/V reuse, and source-background latent composition are shared design elements; we do not claim to introduce these components individually. We will make this relationship explicit and focus the distinction on how source memory is constructed and reused.

FreEdit performs one source-model evaluation at a fixed intermediate noise level, stores the resulting layer-wise attention features, and reuses that same memory throughout target denoising. In the evaluated inversion-free KV-Edit implementation, removing inversion does not remove repeated source evaluations: source features are obtained at successive editing timesteps. Thus, inversion-free editing and single-source-read editing are distinct properties. The comparison below counts source and target sample evaluations separately.

For a closer comparison with inversion-free KV-Edit, we additionally evaluate a FreEdit configuration with more closely aligned mask handling and latent composition, while retaining our single-source-read mechanism. We denote this as FreEdit (comparison configuration).

MethodSource evaluationsTarget evaluationsMean time per image ↓
KV-Edit, inversion-free28289.2604 s
FreEdit (comparison configuration)1273.8406 s

Runtime is averaged over all 620 PIE-Bench examples, following the paper’s timing protocol on an RTX 4090 with 48 GB of VRAM.

We also evaluated editing quality under the existing six-metric protocol:

MethodHPS ×100 ↑BG PSNR ↑BG LPIPS ×1000 ↓BG MSE ×10000 ↓Edit CLIP ×100 ↑IR ×10 ↑
KV-Edit, inversion-free28.069835.18039.41605.087223.01617.6896
FreEdit (comparison configuration)28.411234.926610.61135.620723.46769.1160

FreEdit achieves higher HPS and text-alignment scores with comparable background preservation, while delivering a 2.41× speedup over inversion-free KV-Edit.

2. Isolating QK-CFG (weakness and question 2)

To isolate the contribution of QK-CFG and test the task-specific choices in the paper, we completed four-way ablations—None, Q only, K only, and Q+K—on 280 color edits (40 source images × 7 target colors) and 80 style edits. Within each task, only the Q/K guidance coefficients change; all other settings remain fixed. Each enabled coefficient is 0.8.

QK-CFG guides foreground queries and keys using source features while retaining target values. The semantic experiment evaluates the core source-memory method; the following ablations directly evaluate the additional appearance-guidance component.

Color editing

GuidanceΔE00 ↓Full SSIM ↑BG PSNR ↑
None23.42590.731230.5258
Q only27.41960.779931.8292
K only23.73620.743131.1869
Q + K26.47750.774731.9189

The color ablation supports QK-CFG as a preservation mechanism rather than a direct color-matching mechanism: source guidance helps retain image structure and background, while accurate recoloring benefits from the complementary LCS component. We therefore also compare the combined configuration with the unguided editor:

Color configurationΔE00 ↓Full SSIM ↑BG PSNR ↑
None23.42590.731230.5258
Q + K26.47750.774731.9189
Q + K + LCS14.99530.769030.7857

With LCS, the guided editor improves color accuracy and preservation together over None, showing that the preservation benefit can be retained while achieving the requested color change.

Style editing

GuidanceStyle CLIP ↑Target-text CLIP ↑Aesthetic ↑
None17.541328.63466.1704
Q only18.583728.83836.5219
K only17.604128.56296.3672
Q + K17.945928.43836.5329

The style ablation shows that Q guidance is sufficient to improve both style alignment and aesthetics. Adding K does not further improve alignment, supporting the paper’s Q-only setting rather than enabling both components indiscriminately.

These ablations isolate the benefit of QK-CFG and support its task-specific use: preservation-oriented guidance for color editing and Q-only guidance for style editing. We will include them in the revision.

Thank you again for your constructive feedback. We have added the comparison with inversion-free KV-Edit and controlled QK-CFG ablations to address your main concerns. We would appreciate your reconsideration in light of these results and would be happy to clarify any remaining questions.

Comments on the paper

1 comment

We sincerely thank the Area Chair and the reviewers for their time and constructive feedback.

FreEdit studies a specific efficiency question: how much source computation is needed to guide local image editing? Its single source read constructs fixed attention memory for foreground-only target denoising, without additional training. The rebuttal experiments strengthen the evidence for this design and clarify its relationship to prior work.

We have incorporated the following additions into the revised manuscript:

  1. Direct comparison with inversion-free KV-Edit. For a closer comparison with inversion-free KV-Edit, we evaluate a FreEdit configuration with more closely aligned mask handling and latent composition, retaining the single-source-read mechanism. Under the existing 620-image PIE-Bench protocol, this configuration achieves higher HPS, Edit CLIP, and ImageReward with comparable background preservation, while delivering a 2.41× speedup (3.8406 versus 9.2604 seconds per image) and comparable peak memory. We explain the distinction between a single fixed source read and repeated source evaluations in inversion-free editing. (Addressing 6w7Q W1/Q1 and cAhY's weakness and questions; Appendix H.1, Tables 13–14.)

  2. Controlled QK-CFG ablations. We added None/Q-only/K-only/Q+K comparisons on 280 color edits and 80 style edits, changing only the Q/K coefficients within each task. These experiments isolate QK-CFG's preservation benefit in color editing and support Q-only guidance for style alignment and aesthetics. The additional LCS comparison shows how color correction complements the preservation mechanism. (Addressing 6w7Q W2/Q2; Appendix I, Tables 17–19.)

  3. Quantitative comparison with ReFlex. Using the official FLUX implementation under the same evaluation protocol, FreEdit's paper configuration achieves better results on all six metrics. We clarify the difference between intermediate feature extraction and constructing reusable source memory with one evaluation. (Addressing nySd W1/Q1; Appendix H.2, Table 15.)

  4. Explicit source and target computation. We report source NFE, target NFE, total NFE, and Transformer calls. FreEdit uses one source evaluation and 27 target evaluations, making explicit that “one-step” refers to source-memory construction while target generation remains multi-step. (Addressing nySd W2 and cAhY Q1/Q2; Appendix H.3, Table 16.)

  5. Expanded discussion of efficient editing and inpainting. We added a methodological comparison with ChordEdit, TurboEdit, SwiftEdit, InverFill, and TurboFill, distinguishing source representations, training requirements, backbones, and target sampling budgets. This positions FreEdit alongside complementary approaches to efficient generation. (Addressing nySd W3 and the follow-up recommendation; Appendix J.)

All additions are contained in Appendices H–J and marked in sky blue. We appreciate reviewer cAhY's confirmation that the responses answered most questions and the continued positive recommendation, as well as reviewer nySd's reassessment and decision to raise the score to Weak Accept. We will carry these additions and the expanded discussion into the final version.

We thank the Area Chair and all reviewers again for their consideration and for helping strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.