Pith. sign in

REVIEW 3 major objections 5 minor 41 references

BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation

T0 review · 3 major / 5 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One model can learn both few-step image generation and inversion by estimating average velocity fields in both time directions under a shared physical constraint.

desk verdict Clean MeanFlow extension that jointly trains few-step generation and inversion; useful empirical recipe, soft reversibility claim is the main soft spot. read the letter →

arxiv 2603.24942 v1 pith:ELKPPHQK submitted 2026-03-26 cs.CV

Yasong Dai , Zeeshan Hayder , David Ahmedt-Aristizabal , Hongdong Li This is my paper
classification cs.CV
keywords bidirectionalflowmatchingfew-stepimageeditingaveragevelocityfieldinversiontime-intervalsupervisionconsistencyobjectivediffusionmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Few-step diffusion and flow models are fast, but they invert poorly: large time jumps break the local-linear approximations that ordinary DDIM-style inversion relies on, so edits drift or lose background detail. BiFM trains a single flow-matching network to predict the average velocity of the probability-flow ODE over arbitrary intervals in both directions—noise to image and image to noise—while forcing the two directions to stay negatives of each other through a consistency loss. The shared instantaneous velocity field (taken from a schedule or a pretrained multi-step model) supplies the physical constraint that keeps the large-step predictions on the true trajectory. The result is accurate one-step or few-step inversion and editing without extra inversion networks, and the same model also improves pure generation quality. A sympathetic reader cares because interactive, real-time editing finally becomes reliable under tight sampling budgets.

What carries the argument

The bidirectional extension of the MeanFlow Identity: both forward average velocity u(x_t, t, t′) and backward average velocity u(x_t′, t′, t) are defined from the same instantaneous velocity, and a consistency loss L_BiFM = D(u_θ(x_t,t,t′), −u_θ(x_t′,t′,t)) forces them to remain negatives of each other, enabling reversible large-interval steps.

What would settle it

Train BiFM on a controlled dataset, perform pure one-step inversion of held-out images, and check whether reconstruction error (PSNR/LPIPS) stays competitive with multi-step numerical integration of the teacher ODE; a large gap that grows with interval size would falsify the claimed reversibility.

Watch

Extended reading notes

Core claim

BiFM shows that generation and inversion can be jointly learned inside one flow-matching model by directly estimating bidirectional average velocity fields over continuous time intervals, constrained by a shared instantaneous velocity field and stabilized by a bidirectional consistency objective plus a lightweight time-interval embedding; the resulting model supports accurate one-step inversion and outperforms existing few-step editing and generation methods across standard benchmarks.

Load-bearing premise

That forcing the learned large-interval average velocities to be negatives of each other is enough to keep them on the true probability-flow trajectory even when the time jumps are large.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes BiFM, a flow-matching framework that jointly learns few-step generation and inversion in a single model by predicting average velocity fields over continuous time intervals in both directions. Building on the MeanFlow identity, it defines forward and backward average velocities from a shared instantaneous velocity (predefined schedule or frozen multi-step teacher), adds a bidirectional consistency loss that encourages u_θ(x_t,t,t′) ≈ −u_θ(x_t′,t′,t), and injects a lightweight (t, t′−t) interval embedding into standard backbones (SiT, MMDiT, U-Net). The method is trained from scratch or fine-tuned (LoRA on SD3) and evaluated on image reconstruction, PIE-Bench prompt-guided editing under multi/few/one-step budgets, and generation FID on CIFAR-10, MSCOCO-256, and ImageNet-256.

Significance. If the claims hold, BiFM offers a clean, architecture-agnostic alternative to training-free DDIM-style inversion and to auxiliary inversion networks for few-step editing. Unifying generation and inversion under one average-velocity model with continuous-interval supervision is a useful practical contribution, and the reported gains on reconstruction (Table 2), PIE-Bench (Table 1), and generation (Tables 4–6) would matter for interactive editing and efficient sampling. Strengths include a clear ODE-motivated formulation, seamless integration into popular backbones, and ablations of conditioning, sampling, weighting, and loss norm (Table 3).

major comments (3)
  1. §4.2, Eqs. (9)–(12): The central reversibility claim is that MeanFlow plus the soft consistency term L_BiFM keeps large-interval average velocities on the true PF-ODE in both directions. The paper itself notes that continuous-time reversibility would make the backward average the negative of the forward one along the true trajectory, yet training only softly penalizes the discrepancy with a warm-up weight w(t,t′). In one-step/few-step regimes |t′−t| is large, so residual path inconsistency is not forced to zero. Reconstruction (Table 2) and editing (Table 1) gains are consistent with better inversion but do not isolate whether the learned map remains the integral of the shared instantaneous field versus a convenient bidirectional interpolant. A diagnostic that measures trajectory consistency (e.g., round-trip error under multi-step vs one-step, or residual of the MeanFlow identity on hel
  2. Table 1 and §5.3: Under the few-step (4 NFE) and one-step budgets that motivate the paper, BiFM is not uniformly superior. InstantEdit reports better LPIPS and MSE at 4 NFE; SwiftEdit is competitive on LPIPS at 1 NFE. The narrative of “consistently outperforms existing few-step approaches” should be qualified to the metrics and settings where BiFM actually leads (SSIM/PSNR/CLIP), or additional matched-NFE comparisons and error bars should be provided so the superiority claim is not overstated.
  3. §5.1–5.3 and Algorithms 1–2: Editing experiments fine-tune SD3 with LoRA on MagicBrush and evaluate on PIE-Bench, while generation experiments use different backbones and datasets. It is unclear how much of the editing gain comes from MagicBrush fine-tuning and the extra interval embedding versus the bidirectional consistency objective itself. An ablation that removes L_BiFM (or freezes the interval embedding) on the same SD3 editing setup would make the contribution of the proposed objective load-bearing rather than confounded with domain adaptation.
minor comments (5)
  1. Figure 2 caption and panels (c)–(e): the schematic of DDIM vs BiFM inversion is helpful but dense; labeling the shared instantaneous field v and the two average fields more explicitly would improve readability.
  2. §3.1 time convention: the paper switches between diffusion discrete time and flow-matching t∈[0,1]; a short explicit reparameterization note when citing DDIM equations would reduce ambiguity.
  3. Table 3: the 1-NFE ImageNet ablation FID values (~55) are far from the multi-step SiT numbers in Table 6; stating the exact protocol (epochs, CFG, sample count) in the table caption would avoid confusion.
  4. Supplementary Table A / main Table 1: some InstantEdit NFE entries differ between main text and appendix; align the reported step counts.
  5. Typos and polish: “Amzon” in Acknowledgments; occasional missing spaces around math in the abstract and §4; “direc.” abbreviation in Table 3(a) is unclear.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: BiFM is an empirical training objective (MeanFlow targets + soft bidirectional regularizer), not a derivation that reduces to its inputs by construction.

full rationale

The paper's load-bearing chain is definitional mathematics of average velocity (Eq. 5/6) plus the cited MeanFlow Identity (Eqs. 7–9 from Geng et al.) to obtain regression targets u_tgt from a predefined schedule or frozen pretrained multi-step v, followed by an explicit soft consistency regularizer L_BiFM = D(u_θ(x_t,t,t′), −u_θ(x_t′,t′,t)) (Eqs. 11–12) with warm-up weight. The reverse-average = −forward-average relation is the ordinary integral property of the true trajectory, not a self-definition that forces the learned network or the reported metrics. Targets are external (rectified-flow schedule or pretrained generator); the consistency term is an additional loss, not a tautology. Empirical claims rest on independent benchmarks (PIE-Bench, MSCOCO, CIFAR-10, ImageNet) and ablations, not on fitted constants renamed as predictions. Self-citations (MeanFlow, SD3, REPA, MagicBrush) are ordinary background with non-overlapping authors and do not carry uniqueness theorems that forbid alternatives. No self-definitional loop, no fitted-input-as-prediction, no load-bearing self-citation chain. Score 0 is therefore the correct, proportionate finding.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

The central claim rests on standard flow-matching mathematics, the MeanFlow identity taken from prior work, and several hand-chosen training knobs (consistency weight schedule, time-interval sampler, loss norm, LoRA and embedding design). No new physical entity is postulated; the invented pieces are algorithmic constructs whose only evidence is the paper’s own tables.

free parameters (5)
  • bidirectional consistency weight schedule w(t,t′)
    Hand-designed warm-up profile chosen via ablation (Table 3c); directly scales the load-bearing consistency term in Eq. 12.
  • time-interval sampler (log-normal parameters)
    lognorm(−0.4, 1.0) selected by FID ablation (Table 3b); controls which intervals supervise average velocities.
  • loss norm / Pseudo-Huber parameter p
    Default p=1.0 chosen by ablation (Table 3d); changes residual clipping and final FID.
  • LoRA rank and which MMDiT projections receive adapters
    Follows Flash Diffusion settings but exact ranks/layers are not fully specified; determines capacity of the fine-tuned editor.
  • learning rate and training length (1e−5, 100 epochs SD3; 80 epochs ImageNet)
    Standard optimizer knobs that affect whether the reported metrics are reached.
assumptions (4)
  • domain assumption Flow-matching probability-flow ODE and conditional flow-matching loss are valid generative models for the data distribution.
    Taken as given from Lipman et al. / Liu et al.; entire BiFM construction sits on this ODE (§3.1).
  • domain assumption MeanFlow identity relating average and instantaneous velocity holds for both t<t′ and t>t′.
    Extended from Geng et al. without a new proof; used to define u_tgt in both directions (§4.1–4.2, Eqs. 7–9).
  • ad hoc to paper Average velocity over [t,t′] equals the negative of the average velocity over [t′,t] along the true trajectory, so consistency loss enforces invertibility.
    Physical motivation stated in §4.2; not independently verified beyond reconstruction metrics.
  • domain assumption A pretrained multi-step velocity (or rectified-flow schedule) remains a valid instantaneous target when supervising large time intervals.
    Required for fine-tuning SD3 and for from-scratch rectified-flow training (§4.3).
invented entities (3)
  • BiFM bidirectional average-velocity model
    purpose: Single network that predicts both generation and inversion average velocities under shared instantaneous constraint.
    Core proposed object; evidence is only the paper’s own editing/generation tables.
  • Bidirectional consistency objective L_BiFM
    purpose: Penalize mismatch between forward and negated backward average velocities to stabilize joint training.
    New loss term (Eq. 11); no external falsifiable prediction beyond reported FID/LPIPS.
  • Lightweight time-interval embedding (t, t′−t)
    purpose: Condition backbone on interval start and length for continuous-interval supervision.
    Architectural add-on zero-initialized into SD3/SiT/MMDiT; justified only by ablation FID.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation." pith.science (2026). https://pith.science/paper/ELKPPHQK

@misc{pith2026260324942,
  author       = {Pith},
  title        = {Pith review of: BiFM: Bidirectional Flow Matching for Few-Step Image Editing and Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELKPPHQK}},
  note         = {Machine review of arXiv:2603.24942}
}
abstract

Recent diffusion and flow matching models have demonstrated strong capabilities in image generation and editing by progressively removing noise through iterative sampling. While this enables flexible inversion for semantic-preserving edits, few-step sampling regimes suffer from poor forward process approximation, leading to degraded editing quality. Existing few-step inversion methods often rely on pretrained generators and auxiliary modules, limiting scalability and generalization across different architectures. To address these limitations, we propose BiFM (Bidirectional Flow Matching), a unified framework that jointly learns generation and inversion within a single model. BiFM directly estimates average velocity fields in both ``image $\to$ noise" and ``noise $\to$ image" directions, constrained by a shared instantaneous velocity field derived from either predefined schedules or pretrained multi-step diffusion models. Additionally, BiFM introduces a novel training strategy using continuous time-interval supervision, stabilized by a bidirectional consistency objective and a lightweight time-interval embedding. This bidirectional formulation also enables one-step inversion and can integrate seamlessly into popular diffusion and flow matching backbones. Across diverse image editing and generation tasks, BiFM consistently outperforms existing few-step approaches, achieving superior performance and editability.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 6 linked inside Pith

  1. [1]

    All are worth words: A ViT backbone for diffusion models

    Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A ViT backbone for diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, 2023. 8

    2023

  2. [2]

    Flow map matching with stochastic inter- polants: A mathematical framework for consistency models

    Nicholas Matthew Boffi, Michael Samuel Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic inter- polants: A mathematical framework for consistency models. Transactions on Machine Learning Research, 2025. 2

    2025

  3. [3]

    Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xi- aohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 22560–22570, 2023. 6, 7

    2023

  4. [4]

    Flash diffusion: Accelerating any conditional diffusion model for few steps image generation

    Clement Chadebec, Onur Tasar, Eyal Benaroche, and Ben- jamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. InThe 39th Annual AAAI Conference on Artificial Intelligence, 2025. 5, 6, 11

    2025

  5. [5]

    Image inversion: A survey from gans to diffusion and beyond, 2025

    Yinan Chen, Jiangning Zhang, Yali Bi, Xiaobin Hu, Teng Hu, Zhucun Xue, Ran Yi, Yong Liu, and Ying Tai. Image inversion: A survey from gans to diffusion and beyond, 2025. 3

    2025

  6. [6]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 6

    2009

  7. [7]

    Fireflow: Fast inversion of rec- tified flow for image semantic editing.arXiv preprint arXiv:2412.07517, 2024

    Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rec- tified flow for image semantic editing.arXiv preprint arXiv:2412.07517, 2024. 6

    arXiv 2024

  8. [8]

    Scaling rec- tified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rec- tified flow transformers for high-resolution image synthesis. InForty-first International Conference on Machine Learn- ing, 2024. 5, 11

    2024

Show all 41 references
  1. [9]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. InThe Thirteenth International Conference on Learning Represen- tations, ICLR 2025, Singapore, April 24-28, 2025. OpenRe- view.net, 2025. 2

    2025

  2. [10]

    Renoise: Real image inversion through iterative noising

    Daniel Garibi, Or Patashnik, Andrey V oynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. Renoise: Real image inversion through iterative noising. InEuropean Conference on Computer Vision, pages 395–413. Springer, 2024. 1, 6, 7

    2024

  3. [11]

    Mean flows for one-step genera- tive modeling.arXiv preprint arXiv:2505.13447, 2025

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step genera- tive modeling.arXiv preprint arXiv:2505.13447, 2025. 2, 3, 8

    2025 arXiv

  4. [12]

    Instantedit: Text-guided few-step image editing with piecewise rectified flow

    Yiming Gong, Zhen Zhu, and Minjia Zhang. Instantedit: Text-guided few-step image editing with piecewise rectified flow. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 16808–16817, 2025. 6

    2025

  5. [13]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1, 3, 11

    2020

  6. [14]

    On exact inversion of dpm-solvers

    Seongmin Hong, Kyeonghyun Lee, Suh Yoon Jeon, Hyewon Bae, and Se Young Chun. On exact inversion of dpm-solvers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7069–7078, 2024. 2

    2024

  7. [15]

    Pnp inversion: Boosting diffusion-based editing with 3 lines of code

    Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. InThe Twelfth International Conference on Learning Representations, 2024. 5, 6, 7

    2024

  8. [16]

    Learning multiple layers of features from tiny images, 2009

    Alex Krizhevsky et al. Learning multiple layers of features from tiny images, 2009. 6

    2009

  9. [17]

    Flowedit: Inversion- free text-based editing using pre-trained flow models

    Vladimir Kulikov, Matan Kleiner, Inbar Huberman- Spiegelglas, and Tomer Michaeli. Flowedit: Inversion- free text-based editing using pre-trained flow models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19721–19730, 2025. 6

    2025

  10. [18]

    Bidirectional consistency mod- els

    Liangchen Li and Jiajun He. Bidirectional consistency mod- els. InICML 2024 Workshop on Structured Probabilistic Inference & Generative Modeling, 2024. 3

    2024

  11. [19]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6

    2014

  12. [20]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for genera- tive modeling. InThe Eleventh International Conference on Learning Representations, 2023. 1, 3, 8, 12

    2023

  13. [21]

    Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 1, 3, 4, 8

    2022

  14. [22]

    Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024. 8

    2024 arXiv

  15. [23]

    Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps.Ad- vances in neural information processing systems, 35:5775– 5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps.Ad- vances in neural information processing systems, 35:5775– 5787, 2022. 2

    2022

  16. [24]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 1, 6 9

    2023

  17. [25]

    Swiftedit: Lightning fast text-guided image editing via one-step diffusion

    Trong-Tung Nguyen, Quang Nguyen, Khoi Nguyen, Anh Tran, and Cuong Pham. Swiftedit: Lightning fast text-guided image editing via one-step diffusion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21492–21501, 2025. 2, 6, 7

    2025

  18. [26]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR,

  19. [27]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 8

    2015

  20. [28]

    Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 2

    2022 arXiv

  21. [29]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1, 2, 3, 6, 7, 8

    2010 arXiv

  22. [30]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. InInternational Conference on Learning Represen- tations, 2021. 1, 3

    2021

  23. [31]

    Invertible consistency distillation for text-guided image editing in around 7 steps

    Nikita Starodubcev, Mikhail Khoroshikh, Artem Babenko, and Dmitry Baranchuk. Invertible consistency distillation for text-guided image editing in around 7 steps. InThe Thirty- eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3

    2024

  24. [32]

    Edict: Exact diffusion inversion via coupled transformations

    Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532–22541, 2023. 2

    2023

  25. [33]

    Belm: Bidirec- tional explicit linear multi-step sampler for exact inversion in diffusion models.Advances in Neural Information Process- ing Systems, 37:46118–46159, 2024

    Fangyikang Wang, Hubery Yin, Yue-Jiang Dong, Huminhao Zhu, Hanbin Zhao, Hui Qian, Chen Li, et al. Belm: Bidirec- tional explicit linear multi-step sampler for exact inversion in diffusion models.Advances in Neural Information Process- ing Systems, 37:46118–46159, 2024. 2

    2024

  26. [34]

    Taming rectified flow for inversion and editing

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. InForty-second In- ternational Conference on Machine Learning, 2025. 1, 2, 6

    2025

  27. [35]

    Transition mod- els: Rethinking the generative learning objective, 2025

    Zidong Wang, Yiyuan Zhang, Xiaoyu Yue, Xiangyu Yue, Yangguang Li, Wanli Ouyang, and Lei Bai. Transition mod- els: Rethinking the generative learning objective, 2025. 2

    2025

  28. [36]

    Turboedit: Instant text-based image editing, 2024

    Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing, 2024. 2, 6, 7

    2024

  29. [37]

    Dnaedit: Direct noise alignment for text-guided rectified flow editing.arXiv preprint arXiv:2506.01430,

    Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, and Lei Zhang. Dnaedit: Direct noise alignment for text-guided rectified flow editing.arXiv preprint arXiv:2506.01430,

    arXiv

  30. [38]

    Unveil inversion and invariance in flow transformer for versatile image editing

    Pengcheng Xu, Boyuan Jiang, Xiaobin Hu, Donghao Luo, Qingdong He, Jiangning Zhang, Chengjie Wang, Yunsheng Wu, Charles Ling, and Boyu Wang. Unveil inversion and invariance in flow transformer for versatile image editing. In Proceedings of the Computer Vision and Pattern Recogn...

    2025

  31. [39]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. InThe Thirteenth In- ternational Conference on Learning Representations, 2025. 6, 7

    2025

  32. [40]

    Guoqiang Zhang, J. P. Lewis, and W. Bastiaan Kleijn. Exact diffusion inversion via bi-directional integration approxima- tion, 2023. 2

    2023

  33. [41]

    Magicbrush: A manually annotated dataset for instruction- guided image editing.Advances in Neural Information Pro- cessing Systems, 36:31428–31449, 2023

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction- guided image editing.Advances in Neural Information Pro- cessing Systems, 36:31428–31449, 2023. 6, 11 10 BiFM: Bidirectional Flow Matching for Few-Step Image Editin...

    2023

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.