pith. machine review for the scientific record. sign in

arxiv: 2605.07455 · v1 · submitted 2026-05-08 · 💻 cs.CV

Recognition: no theorem link

EditTransfer++: Toward Faithful and Efficient Visual-Prompt-Guided Image Editing

Lan Chen , Qi Mao , Yiren Song , Yuchao Gu , Siwei Ma
Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords image editingvisual promptsdiffusion modelsedit transfercontrastive refinementcondition compressionhigh-resolution generationin-context learning
0
0 comments X

The pith

EditTransfer++ trains diffusion models on visual example pairs alone by removing text conditioning and reshaping sampling to deliver more faithful and faster image edits.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses failures in visual-prompt-guided image editing where diffusion transformers produce unfaithful results because of text bias and sampling noise. It proposes text-decoupled training that forces the model to extract transformations strictly from image pairs during fine-tuning while keeping optional text use at inference time. A best-worst contrastive mechanism then steers the denoising process toward consistent outputs. Condition compression further cuts token redundancy so that 1024-pixel images can be generated efficiently. Experiments on standard benchmarks and a new EditTransfer-Bench dataset show these changes produce state-of-the-art faithfulness together with large speed gains.

Core claim

EditTransfer++ mitigates textual dominance through text-decoupled training that removes text conditioning during fine-tuning, compelling the model to infer transformations solely from visual evidence. On this visually grounded base, a best-worst contrastive refinement mechanism reshapes denoising trajectories to suppress unfaithful generations and improve seed-to-seed consistency. A condition compression and reuse strategy reduces token redundancy to support efficient 1024-pixel editing.

What carries the argument

Text-decoupled fine-tuning that removes text conditioning combined with best-worst contrastive refinement of denoising trajectories.

If this is right

  • The model supports purely visual prompts at inference while still allowing optional text guidance.
  • Denoising produces more consistent outputs across different random seeds.
  • High-resolution images with a 1024-pixel long edge can be generated with reduced computation.
  • Visual prompt faithfulness exceeds that of prior diffusion-transformer editing methods on existing benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decoupling and compression steps could be tested on other token-heavy diffusion tasks such as video generation or multi-image in-context reasoning.
  • Removing text bias during adaptation might reduce the data needed to teach new visual transformations compared with text-heavy fine-tuning.
  • If the contrastive refinement generalizes, it offers a lightweight way to improve sampling stability in any conditional diffusion model.

Load-bearing premise

Removing text during fine-tuning will make the model learn the intended visual transformations from example pairs without losing necessary capabilities, and contrastive selection will reliably steer sampling away from unfaithful results.

What would settle it

Generate edited images on held-out visual prompt pairs from EditTransfer-Bench where the output fails to reproduce the demonstrated transformation despite clear visual examples, or measure that inference time for 1024-pixel images does not decrease while faithfulness scores remain unchanged.

Figures

Figures reproduced from arXiv: 2605.07455 by Lan Chen, Qi Mao, Siwei Ma, Yiren Song, Yuchao Gu.

Figure 1
Figure 1. Figure 1: Illustration of the edit transfer task. A visual prompt is defined as a pair of images (A, A′ ), where A′ is an edited version of A. Given a query image B, the goal of edit transfer is to apply the transformation demonstrated by (A, A′ ) to B, yielding an edited result B′ . prompt (A, A′ ) to a new query image B, as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Edit transfer results and inference time. Given a visual prompt and a query image, existing methods (a) EditTransfer [9], (b) VisualCloze [10], and (c) RelationAdapter [11] often fail to faithfully reproduce the demonstrated transformation and require long per-image inference time for 1024-long-edge outputs. Our (d) EditTransfer++ more closely follows the visual prompt while achieving much faster inference… view at source ↗
Figure 3
Figure 3. Figure 3: Limitations of na¨ıve DiT-based in-context design for edit transfer. (a) Training with paired text–image supervision causes the model to over￾associate specific visual effects with textual cues, so removing the text greatly weakens the influence of the visual prompt. (b) Even with the same visual prompt and text, the fine-tuned model produces divergent outputs under different random seeds, revealing low vi… view at source ↗
Figure 4
Figure 4. Figure 4: Framework of EditTransfer++. (a) Training pipeline. During training, the text branch is fed with null text to enable text-decoupled learning, while the conditional images (A, A′ , B) are downsampled for condition compression. The full token sequence (A, A′ , B, B′ ) is then processed by the DiT backbone, where causal attention ensures that the conditional tokens (A, A′ , B) remain unaffected by the noisy t… view at source ↗
Figure 5
Figure 5. Figure 5: Detailed illustration of the progressive training procedure. (a) In text-decoupled training, the LoRA modules are first fine-tuned using the standard velocity loss in Eq. (4). (b) In best–worst contrastive refinement, we construct a best–worst contrastive dataset and further update the LoRA with the contrastive objective in Eq. (8) to improve generation consistency. The resulting token sequence [cT ; z; cV… view at source ↗
Figure 6
Figure 6. Figure 6: Data samples in Relation252K and EditTransfer-Bench. (a) In the Relation252K [11] test split, each editing type exhibits nearly identical visual effects across samples, offering limited diversity for evaluating edit generalization. (b) In contrast, EditTransfer-Bench introduces both diverse editing types and varied visual outcomes, enabling a more comprehensive evaluation of edit transfer capabilities. rea… view at source ↗
Figure 7
Figure 7. Figure 7: Feature similarity across timesteps for each image. We extract intermediate features of each image (A, A′ , B, B′ ) at every timestep and compute pairwise similarities over time, visualized as heatmaps. Higher similarity is shown in yellow and lower similarity in dark blue. The conditional images (A, A′ , B) exhibit highly stable features across timesteps, whereas the target B′ changes significantly, suppo… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparisons with ImageBrush [8]. Since ImageBrush is not open-source, we use the examples provided in its original paper for comparison. Given the same visual prompts and query images, our method not only follows the demonstrated transformation more faithfully, but also better preserves the identity and structure of the query image. M 2 × N 2 and the query image B to M 4 × N 4 . The correspondi… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparisons with baselines on Relation252K and EditTransfer-Bench. For each example, given a visual prompt (A, A′ ) and a query image B, we compare the results of EditTransfer [9], VisualCloze [10], RelationAdapter [11], and our EditTransfer++. (a) Results on the Relation252K test split [11]. (b) Results on the proposed EditTransfer-Bench. Across diverse editing types and difficulty levels, ou… view at source ↗
Figure 11
Figure 11. Figure 11: User study results. Bars report the proportion of participants who preferred EditTransfer++ over each baseline (EditTransfer [9], Visual￾Cloze [10], RelationAdapter [11]) in terms of faithfulness, fidelity, and overall preference. Across all three criteria, the majority of users favor our method. • Appearance transfer emphasizes diverse local appear￾ance modifications, such as adding glasses or changing c… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results under different condition-compression set￾tings. All outputs are generated at a 1024-pixel long edge. Overly aggressive downsampling (both (A, A′ ) and B at 256) leads to noticeable degradation, whereas using a 512-pixel visual prompt (A, A′ ) and a 256-pixel query image B offers a better trade-off between visual quality and efficiency. matching the demonstrated transformation. This ad… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative evaluation of EditTransfer++ components. (a) Effect of text-decoupled training (TD); (b) effect of best–worst contrastive refinement (RF); (c) effect of condition compression (DS); (d) effect of condition reuse (RU). Mb denotes the base model without our strategies; M1 = Mb + TD, M2 = M1 + DS, M3 = M2 + RU, and M4 = M3 + RF as the full EditTransfer++ model. Visual Prompt (𝐴, 𝐴′) Query Image 𝐵 … view at source ↗
Figure 14
Figure 14. Figure 14: Generalization of EditTransfer++. Qualitative results on (a) unseen edit variants, (b) unseen edit tasks, and (c) unseen species. In all three settings, our method successfully follows the visual prompt for images beyond the training distribution, indicating strong generalization capability. (condition compression), and M4 = M3 + RU (condition reuse), which is the full model. Impact of text-decoupled trai… view at source ↗
Figure 15
Figure 15. Figure 15 [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
read the original abstract

Visual-prompt-guided edit transfer aims to learn image transformations directly from example pairs, offering more precise and controllable editing than purely text-driven approaches. However, existing diffusion transformer-based methods often fail to faithfully reproduce the demonstrated edits due to structural mismatches between the task and the backbone, including a pretrained bias toward textual conditioning and inherent stochastic instability during sampling. To bridge this gap, we present EditTransfer++, a framework that combines progressively structured training with an efficient conditioning scheme to improve both visual prompt faithfulness and inference efficiency. We first mitigate textual dominance with a text-decoupled training strategy that removes text conditioning during fine-tuning, compelling the model to infer transformations solely from visual evidence while still supporting optional text guidance at inference. On top of this visually grounded model, a best-worst contrastive refinement mechanism reshapes the denoising trajectories to suppress unfaithful generations and improve consistency across random seeds. To alleviate the computational bottleneck of high-resolution in-context editing, we further introduce a condition compression and reuse strategy that reduces token redundancy and enables efficient generation of images with a 1024-pixel long edge. Extensive experiments on existing benchmarks and the proposed EditTransfer-Bench show that EditTransfer++ achieves state-of-the-art visual prompt faithfulness with substantially faster inference than prior methods, suggesting a promising direction for scalable prompt-guided image editing and broader visual in-context learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces EditTransfer++, a diffusion-transformer framework for visual-prompt-guided image editing. It uses text-decoupled fine-tuning to reduce pretrained textual bias and force learning from visual example pairs, a best-worst contrastive refinement step to reshape denoising trajectories toward more faithful outputs, and a condition-compression/reuse strategy to enable efficient 1024-pixel inference. Experiments on existing benchmarks plus the new EditTransfer-Bench report state-of-the-art visual-prompt faithfulness together with substantially faster inference than prior methods.

Significance. If the empirical gains are robust, the work would represent a meaningful engineering advance in controllable, text-light image editing and visual in-context learning. The combination of bias-mitigation training, trajectory reshaping, and token compression directly targets two practical bottlenecks (faithfulness and speed) that have limited deployment of prompt-guided editing at high resolution.

major comments (3)
  1. [Methods (contrastive refinement)] Methods section on best-worst contrastive refinement: the manuscript describes the mechanism at a high level but supplies neither a derivation showing how the contrastive loss alters the reverse-process distribution nor an analysis of its effect on trajectory diversity versus faithfulness. This is load-bearing for the central claim that the reported faithfulness gains on EditTransfer-Bench can be attributed to the proposed refinement rather than to other training choices.
  2. [Training strategy] Training-strategy subsection on text-decoupled fine-tuning: the claim that removing text conditioning 'compels the model to infer transformations solely from visual evidence' is not accompanied by any measurement of capability loss or of how optional text is safely re-introduced at inference without reintroducing textual dominance. Both points are required to justify the faithfulness improvements.
  3. [Experiments] Experimental section and tables: the abstract asserts SOTA faithfulness and speed, yet the provided text contains no quantitative tables, ablation studies, or error analysis that would allow verification of the magnitude of the gains or isolation of each component's contribution.
minor comments (2)
  1. [Abstract / Introduction] The abstract refers to 'progressively structured training' and 'efficient conditioning scheme' without defining these terms; a short paragraph or diagram in the introduction would improve readability.
  2. [Condition compression] Notation for the condition-compression module is introduced without an accompanying equation or pseudocode; adding a compact formulation would aid reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive feedback. We address each major comment point by point below. Where the comments identify areas needing additional detail or clarification, we will revise the manuscript accordingly to strengthen the presentation of our contributions.

read point-by-point responses

  1. Referee: [Methods (contrastive refinement)] Methods section on best-worst contrastive refinement: the manuscript describes the mechanism at a high level but supplies neither a derivation showing how the contrastive loss alters the reverse-process distribution nor an analysis of its effect on trajectory diversity versus faithfulness. This is load-bearing for the central claim that the reported faithfulness gains on EditTransfer-Bench can be attributed to the proposed refinement rather than to other training choices.

    Authors: We agree that a more rigorous derivation and analysis would better substantiate the role of the contrastive refinement. In the revised manuscript we will add a derivation showing how the best-worst contrastive loss modifies the reverse-process distribution, together with quantitative analysis of its effects on trajectory diversity and faithfulness. These additions will help isolate the refinement's contribution from other training choices. revision: yes

  2. Referee: [Training strategy] Training-strategy subsection on text-decoupled fine-tuning: the claim that removing text conditioning 'compels the model to infer transformations solely from visual evidence' is not accompanied by any measurement of capability loss or of how optional text is safely re-introduced at inference without reintroducing textual dominance. Both points are required to justify the faithfulness improvements.

    Authors: We acknowledge the need for explicit measurements. The revised version will include new experiments quantifying any capability loss from text-decoupled fine-tuning and will demonstrate, via controlled ablations, how optional text guidance can be re-introduced at inference time without restoring textual dominance. These results will be presented alongside the existing faithfulness metrics. revision: yes

  3. Referee: [Experiments] Experimental section and tables: the abstract asserts SOTA faithfulness and speed, yet the provided text contains no quantitative tables, ablation studies, or error analysis that would allow verification of the magnitude of the gains or isolation of each component's contribution.

    Authors: The full manuscript contains quantitative tables (Tables 1–3), component ablations (Section 4.3), and error analysis (supplementary material). To improve accessibility, we will expand the main experimental section with additional summary tables and a dedicated ablation table that isolates each component's contribution, ensuring all key results are directly verifiable in the primary text. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical engineering contribution without self-referential derivations or fitted predictions.

full rationale

The paper describes a practical framework combining text-decoupled fine-tuning, best-worst contrastive refinement, and condition compression for visual-prompt-guided editing. No equations, derivations, or parameter-fitting steps are presented that reduce the claimed faithfulness or efficiency gains to quantities defined by the same inputs or self-citations. The abstract and skeptic summary frame the work as an empirical method relying on experimental benchmarks rather than mathematical self-definition or uniqueness theorems imported from prior author work. Central assumptions about overriding text bias and reshaping denoising trajectories are stated as design choices validated by results, not as closed loops where outputs equal inputs by construction. This is a standard non-circular empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is an empirical machine-learning contribution. No mathematical axioms, free parameters, or newly postulated entities are described in the abstract; the central claim rests on the empirical effectiveness of the three proposed engineering changes.

pith-pipeline@v0.9.0 · 5544 in / 1133 out tokens · 28307 ms · 2026-05-11T01:55:04.927514+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 2 internal anchors

  1. [1]

    Prompt-to-prompt image editing with cross attention control,

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y . Pritch, and D. Cohen-Or, “Prompt-to-prompt image editing with cross attention control,” inICLR, 2023

    work page 2023

  2. [2]

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing,

    M. Cao, X. Wang, Z. Qi, Y . Shan, X. Qie, and Y . Zheng, “Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing,” inICCV, 2023

    work page 2023

  3. [3]

    Instructpix2pix: Learning to follow image editing instructions,

    T. Brooks, A. Holynski, and A. A. Efros, “Instructpix2pix: Learning to follow image editing instructions,” inCVPR, 2023

    work page 2023

  4. [4]

    In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer,

    Z. Zhang, J. Xie, Y . Lu, Z. Yang, and Y . Yang, “In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer,” inNeurIPS, 2025

    work page 2025

  5. [5]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    B. F. Labs, S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esseret al., “Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space,”arXiv preprint arXiv:2506.15742, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Visual prompting via image inpainting,

    A. Bar, Y . Gandelsman, T. Darrell, A. Globerson, and A. A. Efros, “Visual prompting via image inpainting,” inNeurIPS, 2022

    work page 2022

  7. [7]

    Images speak in images: A generalist painter for in-context visual learning,

    X. Wang, W. Wang, Y . Cao, C. Shen, and T. Huang, “Images speak in images: A generalist painter for in-context visual learning,” inCVPR, 2023. UNDER REVIEW 13

    work page 2023

  8. [8]

    Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation,

    Y . Yang, H. Peng, Y . Shen, Y . Yang, H. Hu, L. Qiu, H. Koikeet al., “Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation,” inNeurIPS, 2023

    work page 2023

  9. [9]

    Edit transfer: Learning image editing via vision in-context relations,

    L. Chen, Q. Mao, Y . Gu, and M. Z. Shou, “Edit transfer: Learning image editing via vision in-context relations,”arXiv preprint arXiv:2503.13327, 2025

    work page arXiv 2025

  10. [10]

    Visualcloze: A universal image generation framework via visual in-context learning,

    Z.-Y . Li, R. Du, J. Yan, L. Zhuo, Z. Li, P. Gao, Z. Ma, and M.-M. Cheng, “Visualcloze: A universal image generation framework via visual in-context learning,” inICCV, 2025

    work page 2025

  11. [11]

    Relationadapter: Learning and transferring visual relation with diffusion transformers,

    Y . Gong, Y . Song, Y . Li, C. Li, and Y . Zhang, “Relationadapter: Learning and transferring visual relation with diffusion transformers,” inNeurIPS, 2025

    work page 2025

  12. [12]

    Pairedit: Learning semantic variations for exemplar-based image editing,

    H. Lu, J. Chen, Z. Yang, A. T. Gnanha, F. L. Wang, L. Qing, and X. Mao, “Pairedit: Learning semantic variations for exemplar-based image editing,” inNeurIPS, 2025

    work page 2025

  13. [13]

    Personalized vision via visual in-context learning,

    Y . Jiang, Y . Gu, Y . Song, I. Tsang, and M. Z. Shou, “Personalized vision via visual in-context learning,”arXiv preprint arXiv:2509.25172, 2025

    work page arXiv 2025

  14. [14]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inICCV, 2023

    work page 2023

  15. [15]

    Multi-modality cross attention network for image and sentence matching,

    X. Wei, T. Zhang, Y . Li, Y . Zhang, and F. Wu, “Multi-modality cross attention network for image and sentence matching,” inCVPR, 2020

    work page 2020

  16. [16]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeurIPS, 2020

    work page 2020

  17. [17]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inICLR, 2021

    work page 2021

  18. [18]

    Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman, “Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,” inCVPR, 2023

    work page 2023

  19. [19]

    Processpainter: Learning to draw from sequence data,

    Y . Song, S. Huang, C. Yao, H. Ci, X. Ye, J. Liu, Y . Zhang, and M. Z. Shou, “Processpainter: Learning to draw from sequence data,” inSIGGRAPH Asia, 2024

    work page 2024

  20. [20]

    Stable-hair: Real-world hair transfer via diffusion model,

    Y . Zhang, Q. Zhang, Y . Song, J. Zhang, H. Tang, and J. Liu, “Stable-hair: Real-world hair transfer via diffusion model,” inAAAI, 2025

    work page 2025

  21. [21]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inCVPR, 2022

    work page 2022

  22. [22]

    In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024a

    L. Huang, W. Wang, Z.-F. Wu, Y . Shi, H. Dou, C. Liang, Y . Feng, Y . Liu, and J. Zhou, “In-context lora for diffusion transformers,”arXiv preprint arXiv:2410.23775, 2024

    work page arXiv 2024

  23. [23]

    Ominicontrol: Minimal and universal control for diffusion transformer,

    Z. Tan, S. Liu, X. Yang, Q. Xue, and X. Wang, “Ominicontrol: Minimal and universal control for diffusion transformer,” inICCV, 2024

    work page 2024

  24. [24]

    Easycontrol: Adding efficient and flexible control for diffusion transformer,

    Y . Zhang, Y . Yuan, Y . Song, H. Wang, and J. Liu, “Easycontrol: Adding efficient and flexible control for diffusion transformer,” inICCV, 2025

    work page 2025

  25. [25]

    Omniconsistency: Learning style- agnostic consistency from paired stylization data,

    Y . Song, C. Liu, and M. Z. Shou, “Omniconsistency: Learning style- agnostic consistency from paired stylization data,” inNeurIPS, 2025

    work page 2025

  26. [26]

    Makeanything: Harnessing diffusion trans- formers for multi-domain procedural sequence generation.arXiv preprint arXiv:2502.01572, 2025

    Y . Song, C. Liu, and M. Z. Shou, “Makeanything: Harnessing diffusion transformers for multi-domain procedural sequence generation,”arXiv preprint arXiv:2502.01572, 2025

    work page arXiv 2025

  27. [27]

    Photodoodle: Learning artistic image editing from few-shot pairwise data,

    S. Huang, Y . Song, Y . Zhang, H. Guo, X. Wang, M. Z. Shou, and J. Liu, “Photodoodle: Learning artistic image editing from few-shot pairwise data,” inICCV, 2025

    work page 2025

  28. [28]

    Taming rectified flow for inversion and editing,

    J. Wang, J. Pu, Z. Qi, J. Guo, Y . Ma, N. Huang, Y . Chen, X. Li, and Y . Shan, “Taming rectified flow for inversion and editing,” inICML, 2023

    work page 2023

  29. [29]

    Stable flow: Vital layers for training- free image editing,

    O. Avrahami, O. Patashnik, O. Fried, E. Nemchinov, K. Aberman, D. Lischinski, and D. Cohen-Or, “Stable flow: Vital layers for training- free image editing,” inCVPR, 2025

    work page 2025

  30. [30]

    Dit4edit: Diffusion transformer for image editing,

    K. Feng, Y . Ma, B. Wang, C. Qi, H. Chen, Q. Chen, and Z. Wang, “Dit4edit: Diffusion transformer for image editing,” inAAAI, 2025

    work page 2025

  31. [31]

    Instruction-driven multi-weather image translation based on a large-scale image editing model,

    Y . Feng, J. Li, and M. Zhou, “Instruction-driven multi-weather image translation based on a large-scale image editing model,”IEEE TIP, 2025

    work page 2025

  32. [32]

    Consistent image layout editing with diffusion models,

    T. Xia, Y . Zhang, T. Liu, and L. Zhang, “Consistent image layout editing with diffusion models,”IEEE TIP, 2025

    work page 2025

  33. [33]

    Magicbrush: A manually annotated dataset for instruction-guided image editing,

    K. Zhang, L. Mo, W. Chen, H. Sun, and Y . Su, “Magicbrush: A manually annotated dataset for instruction-guided image editing,” inNeurIPS, 2023

    work page 2023

  34. [34]

    Hive: Harnessing human feedback for instructional visual editing,

    S. Zhang, X. Yang, Y . Feng, C. Qin, C.-C. Chen, N. Yu, Z. Chen, H. Wang, S. Savarese, S. Ermonet al., “Hive: Harnessing human feedback for instructional visual editing,” inCVPR, 2024

    work page 2024

  35. [35]

    Image style transfer using convolutional neural networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Image style transfer using convolutional neural networks,” inCVPR, 2016

    work page 2016

  36. [36]

    Cross-image attention for zero-shot appearance transfer,

    Y . Alaluf, D. Garibi, O. Patashnik, H. Averbuch-Elor, and D. Cohen- Or, “Cross-image attention for zero-shot appearance transfer,” inSIG- GRAPH, 2024

    work page 2024

  37. [37]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inICCV, 2017

    work page 2017

  38. [38]

    Attention distillation: A unified approach to visual characteristics transfer,

    Y . Zhou, X. Gao, Z. Chen, and H. Huang, “Attention distillation: A unified approach to visual characteristics transfer,” inCVPR, 2025

    work page 2025

  39. [39]

    Rego: Reference- guided outpainting for scenery image,

    Y . Wang, Y . Wei, X. Qian, L. Zhu, and Y . Yang, “Rego: Reference- guided outpainting for scenery image,”IEEE TIP, 2024

    work page 2024

  40. [40]

    Consistent image inpainting with pre-perception and cross-perception collaborative processes,

    Y . Zhang, Y . Liu, H. Fan, R. Hu, J. Zhang, and Q. Wu, “Consistent image inpainting with pre-perception and cross-perception collaborative processes,”IEEE Transactions on Image Processing, 2025

    work page 2025

  41. [41]

    Anydoor: Zero-shot object-level image customization,

    X. Chen, L. Huang, Y . Liu, Y . Shen, D. Zhao, and H. Zhao, “Anydoor: Zero-shot object-level image customization,” inCVPR, 2024

    work page 2024

  42. [42]

    Paint by example: Exemplar-based image editing with diffusion models,

    B. Yang, S. Gu, B. Zhang, T. Zhang, X. Chen, X. Sun, D. Chen, and F. Wen, “Paint by example: Exemplar-based image editing with diffusion models,” inCVPR, 2023

    work page 2023

  43. [43]

    Specref: A fast training-free baseline of specific reference-condition real image editing,

    S. Chen and J. Huang, “Specref: A fast training-free baseline of specific reference-condition real image editing,” inICICML, 2023

    work page 2023

  44. [44]

    Freeedit: Mask-free reference-based image editing with multi- modal instruction,

    R. He, K. Ma, L. Huang, S. Huang, J. Gao, X. Wei, J. Dai, J. Han, and S. Liu, “Freeedit: Mask-free reference-based image editing with multi- modal instruction,”IEEE TPAMI, 2025

    work page 2025

  45. [45]

    Zero-shot image editing with reference imitation,

    X. Chen, Y . Feng, M. Chen, Y . Wang, S. Zhang, Y . Liu, Y . Shen, and H. Zhao, “Zero-shot image editing with reference imitation,” inNeurIPS, 2025

    work page 2025

  46. [46]

    Pixels: Progressive image xemplar-based editing with latent surgery,

    S. D. Biswas, M. Shreve, X. Li, P. Singhal, and K. Roy, “Pixels: Progressive image xemplar-based editing with latent surgery,” inAAAI, 2025

    work page 2025

  47. [47]

    Poce: Pose- controllable expression editing,

    R. Wu, Y . Yu, F. Zhan, J. Zhang, S. Liao, and S. Lu, “Poce: Pose- controllable expression editing,”IEEE TIP, 2023

    work page 2023

  48. [48]

    Towards understanding cross and self-attention in stable diffusion for text-guided image editing,

    B. Liu, C. Wang, T. Cao, K. Jia, and J. Huang, “Towards understanding cross and self-attention in stable diffusion for text-guided image editing,” inCVPR, 2024

    work page 2024

  49. [49]

    What makes good examples for visual in-context learning?

    Y . Zhang, K. Zhou, and Z. Liu, “What makes good examples for visual in-context learning?” inNeurIPS, 2023

    work page 2023

  50. [50]

    Unifying image processing as visual prompting question answering,

    Y . Liu, X. Chen, X. Ma, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Unifying image processing as visual prompting question answering,” inICML, 2024

    work page 2024

  51. [51]

    Images speak in images: A generalist painter for in-context visual learning,

    X. Wang, W. Wang, Y . Cao, C. Shen, and T. Huang, “Images speak in images: A generalist painter for in-context visual learning,” inCVPR, 2023

    work page 2023

  52. [52]

    B. F. Labs, “Flux,” 2024. [Online]. Available: https://github.com/ black-forest-labs/flux

    work page 2024

  53. [53]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inICLR, 2023

    work page 2023

  54. [54]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inICLR, 2023

    work page 2023

  55. [55]

    Stylegan-nada: Clip-guided domain adaptation of image generators,

    R. Gal, O. Patashnik, H. Maron, A. H. Bermano, G. Chechik, and D. Cohen-Or, “Stylegan-nada: Clip-guided domain adaptation of image generators,”ACM TOG, 2022

    work page 2022

  56. [56]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inICLR, 2019

    work page 2019

  57. [57]

    Splicing vit features for semantic appearance transfer,

    N. Tumanyan, O. Bar-Tal, S. Bagon, and T. Dekel, “Splicing vit features for semantic appearance transfer,” inCVPR, 2022

    work page 2022

  58. [58]

    GPT-4o System Card

    OpenAI, “Gpt-4o system card,” 2024. [Online]. Available: https: //arxiv.org/abs/2410.21276

    work page internal anchor Pith review Pith/arXiv arXiv 2024