Pith. sign in

REVIEW 4 major objections 5 minor 34 references

FlowID reconstructs severely damaged faces for forensic identification by fine-tuning a generative model on one image and masking edits so identity cues survive.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

FlowID combines single-image fine-tuning of Stable Diffusion 3 with attention-derived masks to remove death-related facial artifacts while preserving identity, outperforming open-source editors on the new InjuredFaces proxy benchmark.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Solid engineering of known editing pieces for a real forensic need, plus a usable public proxy benchmark; the transfer claim to next-of-kin recognition is still unbridged. the 4 major comments →

arxiv 2603.29591 v2 pith:4EBYHYEO submitted 2026-03-31 cs.CV

FlowID : Enhancing Forensic Identification with Latent Flow-Matching Models

Jules Ripoll , David Bertoin , Alasdair Newson , Charles Dossal , Jose Pablo Baraybar This is my paper
classification cs.CV
keywords facial reconstructionidentity preservationflow matchingimage editingforensic identificationInjuredFacesattention maskingsingle-image fine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper argues that generative image models can remove graphic trauma from facial photographs of the deceased while keeping enough of the original person for next-of-kin recognition. Manual retouching is slow and fails on severe injuries; FlowID attacks both problems with a short single-image fine-tuning step that adapts a latent flow-matching model to the injured face, plus an attention-derived mask that confines changes to damaged regions so morphological traits, tattoos, and other cues stay intact. The authors also release InjuredFaces, a public benchmark of injured portraits paired with uninjured references and identity embeddings, so methods can be scored on identity preservation rather than generic realism alone. On that benchmark and on ordinary facial attribute edits, FlowID outperforms several open-source editors while remaining light enough for a consumer GPU, which matters because sensitive forensic imagery must stay local. The claimed result is that identity-centric reconstruction can enter real identification workflows without shipping private data off-site.

Core claim

FlowID shows that single-image fine-tuning of a latent flow-matching generator, combined with masks built from the model’s own cross-attention maps, removes severe facial injury artifacts while retaining higher identity similarity and better injury-removal scores than open-source editing baselines, all at a memory cost low enough for local consumer hardware.

What carries the argument

Single-image flow-matching fine-tuning plus attention-based masking: the mask is inferred once from averaged image–concept cross-attention maps, binarized, and used to paste unedited latent regions from the inversion trajectory so only damaged areas change.

Load-bearing premise

Gains measured on public photos of living athletes with sports injuries, scored by face-embedding similarity and a yes/no vision-language injury check, will transfer to real next-of-kin recognition of actual post-mortem forensic images.

What would settle it

A blinded recognition trial in which next-of-kin or forensic staff attempt identification from real post-mortem portraits edited by FlowID versus manual retouching and unedited photos, measuring recognition rate and time-to-decision; no gain for FlowID would falsify the operational claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Medico-legal teams can produce release-ready portraits of the deceased faster than manual retouching for severe trauma.
  • Public missing-person galleries can include reconstructed deceased faces with reduced graphic content while still supporting visual recognition.
  • InjuredFaces gives the community a shared protocol that scores identity preservation, injury removal, and unintended change under extreme damage.
  • Local consumer-GPU inference keeps sensitive post-mortem imagery off external servers.
  • The same fine-tune-plus-mask recipe also improves identity retention on ordinary sequential facial attribute edits.

Where Pith is reading between the lines

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

  • When trauma covers most of the face, attention masks may under-localize; hybrid human-drawn masks would be a natural operational extension.
  • If next-of-kin rely on cues poorly captured by ArcFace embeddings (scars, soft-tissue asymmetries, jewelry wear), table rankings on InjuredFaces may overstate field recognition gains.
  • Parameter-efficient adapters could further cut per-image fine-tuning cost for high-volume morgue pipelines.
  • The same out-of-distribution adaptation pattern may apply to other forensic imaging tasks, such as removing post-mortem scene artifacts while preserving clothing or body marks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes FlowID, an identity-preserving facial reconstruction method for forensic identification of deceased persons with severe facial trauma. It combines single-image fine-tuning of a latent flow-matching model (Stable Diffusion 3) to improve inversion on out-of-distribution injured faces with attention-map-derived binary masks that localize edits and enforce exact reconstruction outside damaged regions (Algorithm 1). The authors introduce InjuredFaces, a proxy benchmark of athlete portraits with sports injuries paired with uninjured identity anchors and ArcFace embeddings, and evaluate identity preservation, injury removal (VLM logits), CLIP faithfulness, FID/CMMD, and LPIPS against UltraEdit, SDEdit, ICEdit, Kontext, and RF-Solver. Results on InjuredFaces and FFHQ attribute edits, plus ablations, show FlowID competitive or leading on the stated metrics while remaining runnable on consumer GPUs; the conclusion asserts operational deployments and successful identifications.

Significance. If the method reliably improves next-of-kin recognition under real post-mortem conditions while remaining privacy-preserving and lightweight, the societal impact is high: forensic and humanitarian identification workflows are bottlenecked by graphic imagery and scarce expert labor. The technical combination of per-image flow-matching fine-tuning and attention-based masking is a clear, fully specified contribution (Alg. 1), and InjuredFaces is a reusable community resource that fills a genuine evaluation gap where real forensic images cannot be shared. Ablations (Table 2) and sequential FFHQ results strengthen the empirical case for the two components. Credit is due for the ethical statement, local-deployment emphasis, and open comparison against recent open-source editors. The main significance risk is that superiority on the athlete-injury proxy may not transfer to true forensic utility.

major comments (4)
  1. [§4.1, Conclusion] §4.1 and Conclusion: The central claim that FlowID “supports identification” and has “contributed to successful identifications” rests on InjuredFaces (living athletes with sports injuries) scored by ArcFace cosine similarity to uninjured references and a binary VLM “is this face injured?” query. Real post-mortem imagery differs systematically (discoloration, bloating, structural trauma, lighting, pose). The manuscript reports no recognition-rate numbers, no blinded next-of-kin or forensic-expert study, and no quantitative bridge between ArcFace/VLM gains and human recognition success. Either provide such evidence or substantially scope the claims to proxy reconstruction quality and treat operational impact as anecdotal.
  2. [§4.1, Table 1(a)] §4.1 evaluation protocol: IP, FID, CMMD, and LPIPS are computed only on samples with VLM score < 0.5, while VLM itself is a primary reported metric and FlowID attains the lowest VLM score (Table 1a). Conditioning the other metrics on the same success filter can favor methods that edit more aggressively or more often pass the VLM threshold, and makes cross-method comparison of identity preservation less transparent. Report unfiltered metrics (or success-rate-stratified metrics) and, ideally, an independent injury-removal judge not used as a filter.
  3. [§4.2, Table 1(a)] §4.2 vs Table 1(a): The text states that FlowID attains “the best CLIP alignment by a wide margin,” but Table 1(a) lists FlowID CLIP = 0.23, which is not the highest (ICEdit 0.25; SDEdit/Kontext/RF-Solver 0.24). This overclaim should be corrected and the discussion of edit faithfulness aligned with the table. Relatedly, the IP margin over Kontext/RF-Solver is small (0.50 vs 0.48); report variance or significance if available.
  4. [§3.2–3.3, Appendix B] §3.2–3.3 and free parameters: Fine-tuning steps N, learning rate, mask timestep t_m, max-pooling overshoot, and inversion strength are free and load-bearing for the editability–reconstruction trade-off (Fig. 2, Fig. 5). The paper should state the exact values used for all Table 1 results (Appendix B is incomplete in the main narrative) and show sensitivity on InjuredFaces, not only qualitative trends, so that the reported outperformance is reproducible under the same protocol.
minor comments (5)
  1. [§4.1, Appendix A] InjuredFaces size is written with corrupted glyphs in the PDF text (“��� injured facial portraits”, “�� ��� distinct identities”, “����� reference images”). State exact counts clearly in §4.1 and Appendix A.
  2. [§3.1–3.3] Notation in §3.1–3.3 is hard to parse (many subscripts rendered as boxes). Clean the probability-flow ODE, CFM loss, and mask equations for camera-ready.
  3. [Figure 3] Figure 3 caption refers to “Figure 3.a” but the figure has no panel labels; align caption and figure.
  4. [§2] Related work could briefly position against classical forensic facial reconstruction / post-mortem imaging literature to help non-CV readers place the contribution.
  5. [§4.1, Conclusion] Clarify whether InjuredFaces will be publicly released with the same license and identity-embedding protocol used in the paper.

Circularity Check

0 steps flagged

No circularity: empirical CV method with external pretrained metrics and no self-referential derivation or fitted-as-prediction steps.

full rationale

FlowID is a purely empirical image-editing pipeline (single-image flow-matching fine-tuning plus attention-map masking) evaluated on a newly constructed proxy benchmark (InjuredFaces) using independent pretrained scorers (ArcFace cosine, CLIP, FID/CMMD, LPIPS, VLM yes/no logits). No equation, theorem, or uniqueness claim is derived; the paper never asserts a first-principles prediction that reduces to its own inputs or to a self-citation. Ablations and comparisons are standard empirical controls, not circular definitions. Operational deployment anecdotes are not quantified as derived results. The work is therefore self-contained against external benchmarks with zero circular steps.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard generative-model machinery plus a handful of free hyperparameters chosen for the forensic setting and the untested transfer assumption that athlete-injury proxies plus embedding/VLM scores predict real next-of-kin recognition. No new physical entities are postulated; the invented objects are the method and the benchmark itself.

free parameters (5)
  • fine-tuning steps N
    Number of single-image gradient steps (illustrated up to 100); chosen by hand to balance inversion quality vs. editability (Fig. 2, Fig. 5).
  • learning rate
    Fixed at 5e-5 with Adam8Bit; no sensitivity analysis reported.
  • mask computation timestep t_m
    Chosen so that the injury concept is localized but before signal vanishes; exact value not ablated beyond the qualitative claim.
  • max-pooling kernel for mask overshoot
    Applied to enlarge the editable region following DiffEdit practice; size not specified or swept.
  • inversion strength / ODE steps
    Controls how far the image is noised; set to retain structure while allowing large edits.
axioms (5)
  • domain assumption ArcFace embeddings of uninjured reference photos form a stable identity anchor whose cosine similarity after editing measures recognition utility.
    Stated in §4.1; used as the primary identity-preservation metric without human validation.
  • domain assumption A binary VLM query (“Is this person’s face injured?”) reliably quantifies injury removal success.
    §4.1; success filter and VLM score both depend on this.
  • domain assumption Cross-attention maps of a DiT, averaged over heads and blocks and binarized by 2-means, correctly localize the textual injury concept.
    §3.3; core of the masking component.
  • domain assumption Single-image flow-matching fine-tuning makes an out-of-distribution injured face a more likely sample under the source prompt, reducing inversion error.
    §3.2 and Fig. 2; foundational to the method.
  • standard math Linear interpolant conditional flow matching and the probability-flow ODE hold for the pretrained SD3 backbone.
    Background §3.1; taken from Lipman et al. / Esser et al.
invented entities (2)
  • InjuredFaces benchmark no independent evidence
    purpose: Provide a shareable proxy dataset of 200 injured athlete faces + uninjured references for identity-preserving reconstruction evaluation.
    Constructed by the authors; no independent public release or external validation yet.
  • FlowID algorithm no independent evidence
    purpose: Combine single-image fine-tuning with attention-mask pasting for forensic facial reconstruction.
    Defined by Alg. 1; the paper’s primary technical contribution.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowID : Enhancing Forensic Identification with Latent Flow-Matching Models." pith.science (2026). https://pith.science/paper/4EBYHYEO

@misc{pith2026260329591,
  author       = {Pith},
  title        = {Pith review of: FlowID : Enhancing Forensic Identification with Latent Flow-Matching Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4EBYHYEO}},
  note         = {Machine review of arXiv:2603.29591}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Every day, many people die under violent circumstances, whether from crimes, war, migration, or climate disasters. Medico-legal and law enforcement institutions document many portraits of the deceased for evidence, but cannot immediately carry out identification on them. While traditional image editing tools can process these photos for public release, the workflow is lengthy and produces suboptimal results. In this work, we leverage advances in image generation models, which can now produce photorealistic human portraits, to introduce FlowID, an identity-preserving facial reconstruction method. Our approach combines single-image fine-tuning, which adapts the generative model to out-of-distribution injured faces, with attention-based masking that localizes edits to damaged regions while preserving identity-critical features. Together, these components enable the removal of artifacts from violent death while retaining sufficient identity information to support identification. To evaluate our method, we introduce InjuredFaces, a novel benchmark for identity-preserving facial reconstruction under severe facial damage. Beyond serving as an evaluation tool for this work, InjuredFaces provides a standardized resource for the community to study and compare methods addressing facial reconstruction in extreme conditions. Experimental results show that FlowID outperforms state-of-the-art open-source methods while maintaining low memory requirements, making it suitable for local deployment without compromising data privacy.

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

34 extracted references · 5 linked inside Pith

  1. [1]

    [Batifolet al., 2025 ] Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Kulal, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv e-prints, pages arXiv–2506,

    2025

  2. [2]

    Ledits++: Limitless im- age editing using text-to-image models

    [Bracket al., 2024 ] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolin´ario Passos. Ledits++: Limitless im- age editing using text-to-image models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861–8870,

    2024

  3. [3]

    Instructpix2pix: Learning to follow image editing instructions

    [Brookset al., 2023 ] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 18392–18402,

    2023

  4. [4]

    Diffedit: Diffusion- based semantic image editing with mask guidance,

    [Couaironet al., 2022 ] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion- based semantic image editing with mask guidance,

    2022

  5. [5]

    Diffusion self- guidance for controllable image generation.Advances in Neural Information Processing Systems, 36:16222–16239,

    [Epsteinet al., 2023 ] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self- guidance for controllable image generation.Advances in Neural Information Processing Systems, 36:16222–16239,

    2023

  6. [6]

    Scaling rectified flow transformers for high-resolution image synthesis

    [Esseret al., 2024 ] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, 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

  7. [7]

    Con- ceptattention: Diffusion transformers learn highly inter- pretable features.ICML 2025: Proceedings of the 42nd International Conference on Machine Learning,

    [Helblinget al., 2025 ] Alec Helbling, Tuna Han Salih Meral, Ben Hoover, Pinar Yanardag, and Duen Horng Chau. Con- ceptattention: Diffusion transformers learn highly inter- pretable features.ICML 2025: Proceedings of the 42nd International Conference on Machine Learning,

    2025

  8. [8]

    Prompt-to-prompt image editing with cross attention con- trol.(2022).URL https://arxiv

    [Hertzet al., 2022 ] Amir Hertz, Ron Mokady, Jay Tenen- baum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention con- trol.(2022).URL https://arxiv. org/abs/2208.01626, 1,

    Pith/arXiv arXiv 2022

  9. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30,

    [Heuselet al., 2017 ] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochre- iter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30,

    2017

  10. [10]

    Denoising diffusion probabilistic models.Advances in neu- ral information processing systems, 33:6840–6851,

    [Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neu- ral information processing systems, 33:6840–6851,

    2020

  11. [11]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3,

    [Huet al., 2022 ] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3,

    2022

  12. [12]

    An edit friendly ddpm noise space: Inversion and manipu- lations

    [Huberman-Spiegelglaset al., 2024 ] Inbar Huberman- Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipu- lations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469–12478,

    2024

  13. [13]

    Tight inversion: Im- age conditioned inversion for real image editing

    [Kadoshet al., 2025 ] Edo Kadosh, Nir Goren, Or Patashnik, Daniel Garibi, and Daniel Cohen-Or. Tight inversion: Im- age conditioned inversion for real image editing. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 6980–6990,

    2025

  14. [14]

    A style-based generator architecture for generative adversarial networks

    [Karraset al., 2019 ] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410,

    2019

  15. [15]

    Imagic: Text-based real image editing with diffusion models

    [Kawaret al., 2023 ] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 6007–6017,

    2023

  16. [16]

    Learning an image editing model without image editing pairs.arXiv preprint arXiv:2510.14978,

    [Kumariet al., 2025 ] Nupur Kumari, Sheng-Yu Wang, Nanx- uan Zhao, Yotam Nitzan, Yuheng Li, Krishna Kumar Singh, Richard Zhang, Eli Shechtman, Jun-Yan Zhu, and Xun Huang. Learning an image editing model without image editing pairs.arXiv preprint arXiv:2510.14978,

    Pith/arXiv arXiv 2025

  17. [17]

    [Labs, 2024] Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux,

    2024

  18. [18]

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

    2023

  19. [19]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    [Liuet al., 2022 ] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    Pith/arXiv arXiv 2022

  20. [20]

    Diffusion hyper- features: Searching through time and space for semantic correspondence

    [Luoet al., 2023 ] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyper- features: Searching through time and space for semantic correspondence. InThirty-seventh Conference on Neural Information Processing Systems,

    2023

  21. [21]

    SDEdit: Guided image synthesis and editing with stochas- tic differential equations

    [Menget al., 2022 ] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochas- tic differential equations. InInternational Conference on Learning Representations,

    2022

  22. [22]

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

    [Mokadyet al., 2023 ] Ron Mokady, Amir Hertz, Kfir Aber- man, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6038–6047,

    2023

  23. [23]

    Zero-shot image-to-image translation

    [Parmaret al., 2023 ] 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

  24. [24]

    Scalable diffusion models with transformers

    [Peebles and Xie, 2023] William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 4195–4205, October

    2023

  25. [25]

    Emu edit: Precise image edit- ing via recognition and generation tasks

    [Sheyninet al., 2024 ] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image edit- ing via recognition and generation tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871–8879,

    2024

  26. [26]

    Deep unsupervised learning using nonequilibrium thermodynam- ics

    [Sohl-Dicksteinet al., 2015 ] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynam- ics. InInternational conference on machine learning, pages 2256–2265. pmlr,

    2015

  27. [27]

    Denoising diffusion implicit models

    [Songet al., 2021a ] Jiaming Song, Chenlin Meng, and Ste- fano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,

    2021

  28. [28]

    Plug-and-play diffusion fea- tures for text-driven image-to-image translation

    [Tumanyanet al., 2023 ] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion fea- tures for text-driven image-to-image translation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921–1930, June

    2023

  29. [29]

    Unitune: Text-driven image editing by fine tuning a diffu- sion model on a single image.ACM Trans

    [Valevskiet al., 2023] Dani Valevski, Matan Kalman, Eyal Molad, Eyal Segalis, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning a diffu- sion model on a single image.ACM Trans. Graph., 42(4), July

    2023

  30. [30]

    Taming rectified flow for inversion and editing

    [Wanget al., 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,

    Pith/arXiv arXiv 2024

  31. [31]

    Cmmd: Contrastive multi-modal diffusion for video-audio conditional modeling

    [Yanget al., 2024 ] Ruihan Yang, Hannes Gamper, and Sebas- tian Braun. Cmmd: Contrastive multi-modal diffusion for video-audio conditional modeling. InEuropean Conference on Computer Vision, pages 214–226. Springer,

    2024

  32. [32]

    The unreasonable effectiveness of deep features as a perceptual metric

    [Zhanget al., 2018 ] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595,

    2018

  33. [33]

    In-context edit: Enabling instructional image editing with in-context generation in large scale dif- fusion transformer.arXiv preprint arXiv:2504.20690,

    [Zhanget al., 2025 ] Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. In-context edit: Enabling instructional image editing with in-context generation in large scale dif- fusion transformer.arXiv preprint arXiv:2504.20690,

    Pith/arXiv arXiv 2025

  34. [34]

    Ultraedit: Instruction- based fine-grained image editing at scale.Advances in Neural Information Processing Systems, 37:3058–3093,

    [Zhaoet al., 2024 ] Haozhe Zhao, Xiaojian Shawn Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction- based fine-grained image editing at scale.Advances in Neural Information Processing Systems, 37:3058–3093,

    2024

This paper was first reviewed by grok-4.5 on July 13, 2026.