Pith. sign in

REVIEW 2 major objections 5 minor 3 cited by

DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DNAEdit claims directly refining Gaussian noise removes accumulated drift in rectified-flow inversion, yielding the best reported reconstruction and editing quality among training-free methods.

desk verdict Promising new inversion idea for flow editing, but the central derivation in Eq. (6) has a sign inconsistency that makes the claimed fixed point false as written; worth a serious referee, not acceptance yet. read the letter →

arxiv 2506.01430 v1 pith:R2Z54BJD submitted 2025-06-02 cs.CV

Chenxi Xie , Minghan Li , Shuai Li , Yuhui Wu , Qiaosi Yi , Lei Zhang This is my paper
classification cs.CV
keywords rectifiedflowtraining-freeimageeditingnoiseinversionvelocityfieldalignmenttext-guidedPIE-Benchlong-promptbenchmark
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

Rectified-flow image editing usually reconstructs a source image by reversing the generation trajectory, approximating each noisy latent from the previous one and thereby accumulating drift. The paper proposes Direct Noise Alignment (DNA): instead of re-noising the image, start from a random Gaussian noise and iteratively move it toward a noise that, when fed into the pretrained flow model, reproduces the image. The claim is that this noise-domain feedback eliminates the accumulated error of inversion-based methods while keeping the method training-free and model-agnostic. With a companion Mobile Velocity Guidance (MVG) that blends source- and target-conditioned velocities, the resulting DNAEdit method is reported to achieve the best reconstruction accuracy and the best average ranking on PIE-Bench and on a new long-prompt benchmark DNA-Bench.

What carries the argument

The central object is the velocity gap $\Delta v^{\mathrm{DNA}}_t$ between the linear interpolation velocity and the velocity predicted by the pretrained rectified-flow model. DNA uses this difference as a feedback signal to shift the Gaussian noise sample $S_t$ at every timestep; the paper argues that because rectified-flow latents are defined by direct interpolation between noise and image, correcting the noise in noise-space avoids the accumulated approximation error of previously estimating $Z_t$ from $Z_{t+1}$. The companion object is the residual offset $\Delta x^{\mathrm{DNA}}_t$, which aligns the re-denosing velocity with the inversion trajectory and thereby preserves unedited regions.

What would settle it

Evaluate the fixed-point identity after Eq. (6): substituting $S_t = S_{t+1} + (v^{\mathrm{linear}}_t - v^{\mathrm{src}}_t)\sigma_{t+1}$ and $v^{\mathrm{linear}}_t = (S_{t+1}-Z_{t+1})/\sigma_{t+1}$ into $(S_t-Z_{t+1})/\sigma_{t+1}$ yields $2v^{\mathrm{linear}}_t - v^{\mathrm{src}}_t$, not $v^{\mathrm{src}}_t$; equality holds only when $v^{\mathrm{linear}}_t = v^{\mathrm{src}}_t$. Running Algorithm 1 with the noise update sign reversed and checking whether the reconstruction MSE remains near $0.010$ would settle whether the claimed convergence depends on a hidden sign convention.

Watch

Extended reading notes

Core claim

At every timestep $t$, DNA constructs an interpolated latent $Z^*_t$ from the current noise $S_{t+1}$ and the reference latent $Z_{t+1}$, evaluates the source-conditioned velocity $v^{\mathrm{src}}_t = v_\theta(Z^*_t, \psi_{\mathrm{src}})$, and compares it with the linear velocity $v^{\mathrm{linear}}_t = (S_{t+1}-Z_{t+1})/\sigma_{t+1}$. The difference, $\Delta v^{\mathrm{DNA}}_t = v^{\mathrm{linear}}_t - v^{\mathrm{src}}_t$, is fed back into the noise: $S_t = S_{t+1} + \Delta v^{\mathrm{DNA}}_t \sigma_{t+1}$. Iterating from large to small timesteps yields a structured noise that, combined with stored residual offsets $\Delta x^{\mathrm{DNA}}_t = Z^*_t - Z_t$, allows near-exact reconstruction and supports target-prompt editing with a moving reference velocity. The paper reports reconstruction MSE of $0.010$ and SSIM of $0.830$, versus $0.015$ and $0.726$ for FireFlow, and an average rank of $2.50$ on PIE-Bench.

Load-bearing premise

The whole approach depends on the claim that after moving the noise, the straight-line path from noise to image points exactly where the model predicts; the equations as written only make that true if the predicted velocity is implicitly negated, and that convention is not stated.

Editorial extensions

If this is right

  • DNA inversion attains the best reconstruction among the compared rectified-flow inversion methods, with MSE 0.010, LPIPS 0.110 and SSIM 0.830, under similar numbers of function evaluations to prior methods.
  • DNAEdit ranks first on average among text-guided editing methods on PIE-Bench (average rank 2.50) while using either FLUX or SD3.5 backbones.
  • The method is training-free and model-agnostic, so it can be applied to other rectified-flow generators without modifying attention or adapter layers.
  • Long prompts on DNA-Bench improve background preservation for inversion-based methods, and DNAEdit remains the best-ranked method on that benchmark.
  • Adding Mobile Velocity Guidance with a moving reference balances fidelity and editability: structure distance drops from 33.98 to 18.87 while edited-region CLIP similarity is nearly unchanged.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because DNA stores the residual offsets from inversion, the same inverted noise could be reused for several target prompts without recomputing the inversion; the paper's experiments all re-run DNA per edit, so this is an unstated efficiency gain.
  • The paper's theoretical connection to FlowEdit suggests a natural stress test: run DNAEdit on tasks outside the pretrained model's prior, where the paper's own limitations section says it may fail, and observe where reconstruction MSE breaks down to map the method's coverage.
  • DNA-Bench's construction uses long source and target prompts; a direct follow-up would vary prompt length continuously and measure PSNR and CLIP similarity, isolating how much of the reported gain comes from the benchmark design rather than the editing algorithm.
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

2 major / 5 minor

Summary. The paper proposes DNAEdit, a training-free editing pipeline for rectified-flow text-to-image models. DNA iteratively updates a Gaussian noise sample by comparing the velocity of the linear interpolation between the current noise and the source image with the model's predicted source-conditioned velocity, aiming to obtain a 'structured noise' that reconstructs the image with low drift. A second module, MVG, blends the target-conditioned velocity with a moving-reference velocity during re-denosing to balance fidelity and editability. The authors also introduce DNA-Bench, a long-prompt benchmark, and report state-of-the-art reconstruction and editing results on PIE-Bench and DNA-Bench (Tables 1-3).

Significance. If the mechanism works as claimed, DNA is a conceptually simple and model-agnostic alternative to higher-order RF inversion methods, and the reported gains are substantial: Table 1 shows MSE 0.010 and SSIM 0.830 versus FireFlow's 0.015 and 0.726, and Table 2 shows the best average rank (2.50) on PIE-Bench. The paper includes full-method ablations (Table 4), a new long-prompt benchmark, and a clear presentation of the intended algorithm. The main barrier is a sign inconsistency in the central update: the derivation after Eq. (6) and Algorithm 1 do not have the fixed point claimed in the text, and the appendix uses a different sign convention. This must be resolved before the theoretical and algorithmic claims can be accepted.

major comments (2)
  1. [Section 3.2, Eq. (6), Algorithm 1, and Appendix A.1] Under the RF objective in Eq. (2), v_theta is trained to predict Z_1 - Z_0, i.e., the noise-to-image velocity. With the interpolation in Eq. (5), the linear velocity v_linear = (S_{t+1} - Z_{t+1})/sigma_{t+1} points from the image toward the noise, so at the intended aligned state we should have v_linear = -v_src, not v_linear = v_src. Substituting the stated update S_t = S_{t+1} + (v_linear - v_src)*sigma_{t+1} into the quantity (S_t - Z_{t+1})/sigma_{t+1} gives 2*v_linear - v_src, which equals v_src only when v_linear = v_src. The sentence after Eq. (6) claiming that the updated linear velocity 'matches the predicted velocity' is therefore not derivable from the displayed equations. Appendix A.1 uses a different definition, Delta_v_DNA = (Z_{t+1} - Z*_t)/(sigma_{t+1} - sigma_t) - v_src = -v_linear - v_src, which is sign-consistent with Eq. (2). The main text, Algorithm 1, and Algorithms 3-4 must be reconciled with one clear convention; as written, the DNA loop does not have the claimed fixed point, and the empirical results in Tables 1-3 do not by themselves tell the reader which sign was actually implemented.
  2. [Appendix B, Eq. (18)] The proof of the parallelogram rule for DNAEdit is incomplete. In the passage from the expansion of Z_tgt_t - Z_src_t to the line involving the sum of (v_tgt_i - v_src_i)*(sigma_{i+1} - sigma_i), the proof assumes Z_tgt_0 = Z_src_0 and identifies successive latent differences with the stored velocities; these identities are asserted rather than derived from the DNA/MVG update equations. Since this appendix is the paper's theoretical justification for the claimed equivalence with FlowEdit, it needs a rigorous derivation or the claim should be softened.
minor comments (5)
  1. [Algorithm 1] The line 'Delta_x_DNA_t = Z*_t - Z*' contains an undefined symbol Z*; from Eq. (7) and the surrounding text it should read Z*_t - Z_t.
  2. [Section 3.2, after Eq. (7)] The phrase 'sigma_{t+1}: 1 -> 0' conflicts with the forward/reverse ordering described in Eq. (3); please define the monotonicity of sigma_t explicitly so that the loop direction in Algorithm 1 can be verified.
  3. [Table 2 and Section 4.1] The FTEdit row uses numbers taken from the FTEdit paper rather than from the shared experimental environment; please mark this clearly in the table caption and discuss the caveat in the text.
  4. [Appendix A.1, Eq. (11)] Equation (11) defines Delta_v_DNA_t differently from Eq. (6) in the main text; after fixing the sign, unify the notation across the main paper and the appendix.
  5. [Figure 2(a)] The arrows for v_linear and v_src should be annotated with the chosen sign convention to prevent misreading; the current figure is ambiguous about which direction corresponds to the model's predicted velocity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DNAEdit's core reconstruction and editing claims are validated on external benchmarks; the noise-alignment procedure is a defined, training-free algorithm rather than a fitted prediction or an imported self-citation.

full rationale

I walked the derivation chain in Sec. 3.2 (Eqs. 5-7, Algorithm 1) and the appendix analyses. The method defines the noise update explicitly in terms of v_linear_t = (S_{t+1} - Z_{t+1}) / sigma_{t+1} and v_src_t = v_theta(Z*_t, psi_src), both computed from the pretrained external RF model and the current interpolated latent; no parameter is fitted to the PIE-Bench or DNA-Bench outcomes before the headline numbers are produced. Table 1 and Tables 2-3 compare against external benchmarks (PIE-Bench, and a GPT-4o-extended benchmark called DNA-Bench), so the central claims are not constructed from the evaluation data. The MVG coefficient eta = 0.8 is chosen by an ablation (Appendix E.2) and is presented as a user trade-off hyperparameter, not as a predicted scientific quantity; it does not reduce the core claim that noise alignment reduces drift to a fit. The only self-citation is reference [10], used in Appendix F as context for a video-editing experiment; it is not load-bearing for any main result, and there is no imported uniqueness theorem or prior-work ansatz forcing the method choice. Appendix B's equivalence between DNAEdit and FlowEdit is a consistency and interpretation argument, and since it is not the source of the empirical claims, it does not create a self-referential dependency. One caveat is outside circularity: the sentence after Eq. (6) asserting that the updated linear velocity matches v_theta may be algebraically inconsistent with the displayed sign conventions, but an algebraic sign error is a correctness issue, not a reduction of the result to its inputs, so it does not raise the circularity score under the stated rubric.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The ledger shows that the central claim buys its power from the pretrained velocity fields (FLUX, SD3.5) and from an update rule whose fixed point is asserted rather than derived. The only invented algorithmic objects are the structured noise and the residual offset; no physical entities are postulated. Free parameters are confined to guidance and step hyperparameters, with eta selected on the evaluation benchmarks.

free parameters (4)
  • MVG coefficient eta = 0.8 (both FLUX and SD3.5 versions)
    Blends target-conditioned and mobile-reference velocities (Eq. 10); chosen via visual ablation over 0.7-1.0 (Appendix E.2) on the same task family used for the headline tables; no held-out split described.
  • MVG start step t_s = 4 (FLUX), 13 (SD3.5)
    Start of the guidance phase; set by hand to avoid excessive early changes (Section 4.1, Appendix C.1).
  • CFG scale = 2.5 (FLUX), 3.5 (SD3.5)
    Guidance weight used in DNA and MVG velocity evaluations; set per model with no stated tuning protocol.
  • DNA steps T = 28 (FLUX), 40 (SD3.5)
    Number of alignment steps; affects reported NFE (56) and the drift-error tradeoff; set by hand.
assumptions (4)
  • domain assumption The pretrained RF velocity field v_theta is accurate enough on interpolated noisy latents that aligning predicted and linear velocities yields a straight, low-error trajectory.
    Core assumption of DNA; the model priors (FLUX, SD3.5) are trusted to provide reliable v_src. Invoked in Section 3.2 (Eq. 5-6) and bounded by the authors' own Limitations paragraph: edits outside the base model's prior fail.
  • domain assumption The velocity field is smooth over the latent displacement Z_t - Z*_t, so v_theta(Z_t) is approximately v_theta(Z*_t).
    Used to justify that the updated latent keeps the aligned velocity ('Therefore, it can be deduced...' in Section 3.2 after Eq. (7)); no Lipschitz bound or empirical check is given.
  • ad hoc to paper The noise update S_t = S_{t+1} + (v_linear - v_src)*sigma_{t+1} moves S toward the ideal noise and has the claimed fixed point.
    The paper asserts the updated linear velocity matches v_src, which does not follow from its own equations (substitution gives 2*v_linear - v_src); convergence is not analyzed. The central mechanism depends on this update.
  • domain assumption GPT-4o-generated long prompts faithfully expand PIE-Bench short prompts without altering editing intent.
    DNA-Bench quality depends on the MLLM expansion preserving the editing target; only a few examples are shown (Appendix C.2, Fig. 11) and no human validation is reported.
invented entities (2)
  • Structured (ideal) Gaussian noise S*
    purpose: Target noise for DNA: a noise that, interpolated with the reference image, follows the model's velocity field, giving drift-free reconstruction.
    Defined only through the alignment fixed point of the DNA iteration; the only observable handle is downstream reconstruction quality, which is measured after the fact.
  • Residual offset delta_x_DNA_t independent evidence
    purpose: Per-timestep latent correction added back during re-denosing to keep editing latents aligned with inversion latents.
    Ablation Exp. 3 versus Exp. 1 (Table 4) shows structure distance improving from 31.90 to 24.93 and PSNR from 22.61 to 23.75 dB, an in-paper falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing." pith.science (2026). https://pith.science/paper/R2Z54BJD

@misc{pith2026250601430,
  author       = {Pith},
  title        = {Pith review of: DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2Z54BJD}},
  note         = {Machine review of arXiv:2506.01430}
}
read the original abstract

Leveraging the powerful generation capability of large-scale pretrained text-to-image models, training-free methods have demonstrated impressive image editing results. Conventional diffusion-based methods, as well as recent rectified flow (RF)-based methods, typically reverse synthesis trajectories by gradually adding noise to clean images, during which the noisy latent at the current timestep is used to approximate that at the next timesteps, introducing accumulated drift and degrading reconstruction accuracy. Considering the fact that in RF the noisy latent is estimated through direct interpolation between Gaussian noises and clean images at each timestep, we propose Direct Noise Alignment (DNA), which directly refines the desired Gaussian noise in the noise domain, significantly reducing the error accumulation in previous methods. Specifically, DNA estimates the velocity field of the interpolated noised latent at each timestep and adjusts the Gaussian noise by computing the difference between the predicted and expected velocity field. We validate the effectiveness of DNA and reveal its relationship with existing RF-based inversion methods. Additionally, we introduce a Mobile Velocity Guidance (MVG) to control the target prompt-guided generation process, balancing image background preservation and target object editability. DNA and MVG collectively constitute our proposed method, namely DNAEdit. Finally, we introduce DNA-Bench, a long-prompt benchmark, to evaluate the performance of advanced image editing models. Experimental results demonstrate that our DNAEdit achieves superior performance to state-of-the-art text-guided editing methods. Codes and benchmark will be available at \href{ https://xiechenxi99.github.io/DNAEdit/}{https://xiechenxi99.github.io/DNAEdit/}.

Figures

Figures reproduced from arXiv: 2506.01430 by the authors.

Figure 1
Figure 1. Illustration of (a) existing re-noise based RF inversion methods and (b) our DNA. The ideal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of (a) DNA and (b) MVG, which collectively build our DNAEdit algorithm. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison on reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of different text-guided image editing methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison between editing re￾sults using short and long prompt. Results on DNA-Bench [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Direct noise alignment (DNA) for timestep t. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: More visualization of the difference in Gaussian noise between [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: DNA reconstruction and editing without and with residual offset. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Illustration of the equivalence between FlowEdit and DNAEdit. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Parallelogram rule in FlowEdit and DNAEdit. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Detailed prompt for constructing DNA-Bench and samples from DNA-Bench. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: More results of DNAEdit on real images. Original(Left) [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: More results of DNAEdit on real images. Original(Left) [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: More visual comparisons on PIE-Bench. Left: FLUX-based, Right: SD-based. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: More visual comparisons on DNA-Bench. Left: FLUX-based, Right: SD-based. [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Visual comparison between results from PIE-Bench and DNA-Bench. [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]
Figure 17
Figure 17. Figure 17: Comparison between different MVG coefficients [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]
Figure 18
Figure 18. Figure 18: Visual results of DNAEdit for video editing task. [PITH_FULL_IMAGE:figures/full_fig_p022_18.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A single flow-matching model can learn bidirectional average velocities under a shared instantaneous field and a consistency loss, improving few-step image editing and generation over prior few-step baselines.

  2. Fine-structure Preserved Real-world Image Super-resolution via Transfer VAE Training

    cs.CV 2025-07 conditional novelty 6.0 of 10

    A transfer training scheme converts Stable Diffusion's 8x VAE into a 4x VAE that stays compatible with the pretrained UNet, improving fine-structure preservation in real-world super-resolution at lower FLOPs.

  3. Efficient Difficulty-Aware Dynamic Routing for Diffusion-Based Real-World Image Super-Resolution

    cs.CV 2026-07 reject novelty 4.0 of 10

    DDR-SR routes each real-world low-resolution image to one of two diffusion experts based on a high-frequency-loss difficulty score, using a low-compression VAE for hard images and a high-compression VAE for easy image...

Reference graph

Works this paper leans on

33 extracted references · 14 canonical work pages · cited by 3 Pith papers

  1. [1]

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

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023

    work page 2023

  2. [2]

    Fireflow: Fast inversion of rectified 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 rectified flow for image semantic editing.arXiv preprint arXiv:2412.07517, 2024. 22

    arXiv 2024

  3. [3]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first International Conference on Machine Learning, 2024

    2024

  4. [4]

    Instruct- nerf2nerf: Editing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct- nerf2nerf: Editing 3d scenes with instructions. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023

    work page 2023

  5. [5]

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

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to- prompt image editing with cross-attention control. InThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  6. [6]

    Direct inversion: Boosting diffusion- based editing with 3 lines of code.arXiv preprint arXiv:2310.01506, 2023

    Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion- based editing with 3 lines of code.arXiv preprint arXiv:2310.01506, 2023

    arXiv 2023

  7. [7]

    Flowedit: Inversion- free text-based editing using pre-trained flow models.arXiv preprint arXiv:2412.08629, 2024

    Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion- free text-based editing using pre-trained flow models.arXiv preprint arXiv:2412.08629, 2024

    arXiv 2024

  8. [8]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux , 2024. Accessed: 2024- 11-14

    work page 2024

Show all 33 references
  1. [9]

    Official weights of FLUX.1 dev.https://huggingface.co/black-forest-labs/ FLUX.1-dev, 2024

    Black Forest Labs. Official weights of FLUX.1 dev.https://huggingface.co/black-forest-labs/ FLUX.1-dev, 2024. Accessed: 2024-11-14

    2024

  2. [10]

    Five: A fine-grained video editing benchmark for evaluating emerging diffusion and rectified flow models.arXiv preprint arXiv:2503.13684, 2025

    Minghan Li, Chenxi Xie, Yichen Wu, Lei Zhang, and Mengyu Wang. Five: A fine-grained video editing benchmark for evaluating emerging diffusion and rectified flow models.arXiv preprint arXiv:2503.13684, 2025

    2025 arXiv

  3. [11]

    Zone: Zero-shot instruction-guided local editing

    Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. Zone: Zero-shot instruction-guided local editing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6254–6263, 2024

    2024

  4. [12]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  5. [13]

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

    Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  6. [14]

    Adams bashforth moulton solver for inversion and editing in rectified flow.arXiv preprint arXiv:2503.16522, 2025

    Yongjia Ma, Donglin Di, Xuan Liu, Xiaokai Chen, Lei Fan, Wei Chen, and Tonghua Su. Adams bashforth moulton solver for inversion and editing in rectified flow.arXiv preprint arXiv:2503.16522, 2025

    2025

  7. [15]

    Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models.arXiv preprint arXiv:2305.16807, 2023

    Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models.arXiv preprint arXiv:2305.16807, 2023

    2023 arXiv

  8. [16]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023

    2023

  9. [17]

    Zero-shot image-to-image translation

    Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. InACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023

    2023

  10. [18]

    Semantic image inversion and editing using rectified stochastic differential equations.arXiv preprint arXiv:2410.10792, 2024

    Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations.arXiv preprint arXiv:2410.10792, 2024

    2024 arXiv

  11. [19]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021

    2021

  12. [20]

    Official weights of SD3 medium diffusers

    StablityAI. Official weights of SD3 medium diffusers. https://huggingface.co/stabilityai/ stable-diffusion-3-medium-diffusers, 2024. Accessed: 2024-11-14

    2024

  13. [21]

    Official weights of SD3.5 medium

    StablityAI. Official weights of SD3.5 medium. https://huggingface.co/stabilityai/ stable-diffusion-3.5-medium, 2024. Accessed: 2024-11-14

    2024

  14. [22]

    Gpt-4o system card, 2024

    OpenAI Team. Gpt-4o system card, 2024

    2024

  15. [23]

    Plug-and-play diffusion features for text- driven image-to-image translation

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text- driven image-to-image translation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921–1930, 2023. 23

    1921

  16. [24]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    2025 arXiv

  17. [25]

    Taming rectified flow for inversion and editing.arXiv preprint arXiv:2411.04746, 2024

    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.arXiv preprint arXiv:2411.04746, 2024

    2024 arXiv

  18. [26]

    Wang, A.C

    Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing, page 600–612, 2004

    2004

  19. [27]

    Godiva: Generating open-domain videos from natural descriptions

    Chengdong Wu, Ling-Qiao Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. 2021

    2021

  20. [28]

    Unveil inversion and invariance in flow transformer for versatile image editing.arXiv preprint arXiv:2411.15843, 2024

    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.arXiv preprint arXiv:2411.15843, 2024

    2024 arXiv

  21. [29]

    Inversion-free image editing with language-guided diffusion models

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with language-guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9452–9461, 2024

    2024

  22. [30]

    Eedit: Rethinking the spatial and temporal redundancy for efficient image editing.arXiv preprint arXiv:2503.10270, 2025

    Zexuan Yan, Yue Ma, Chang Zou, Wenteng Chen, Qifeng Chen, and Linfeng Zhang. Eedit: Rethinking the spatial and temporal redundancy for efficient image editing.arXiv preprint arXiv:2503.10270, 2025

    2025

  23. [31]

    Object-aware inversion and reassembly for image editing.arXiv preprint arXiv:2310.12149, 2023

    Zhen Yang, Ganggui Ding, Wen Wang, Hao Chen, Bohan Zhuang, and Chunhua Shen. Object-aware inversion and reassembly for image editing.arXiv preprint arXiv:2310.12149, 2023

    2023 arXiv

  24. [32]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018

    2018

  25. [33]

    Kv-edit: Training-free image editing for precise background preservation.arXiv preprint arXiv:2502.17363, 2025

    Tianrui Zhu, Shiyi Zhang, Jiawei Shao, and Yansong Tang. Kv-edit: Training-free image editing for precise background preservation.arXiv preprint arXiv:2502.17363, 2025. 24

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.