Pith. sign in

REVIEW 4 major objections 5 minor 67 references

UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read UniVVT claims video virtual try-on can be done end-to-end with no masks, poses, or warping, by having an MLLM read the scene and garment into compact task tokens that guide a video diffusion transformer.

desk verdict A genuinely new maskless try-on architecture whose training-data foundation is the soft spot: synthetic sources are manually filtered with no retention stats or distribution-shift checks. read the letter →

arxiv 2608.05745 v1 pith:OWNM4HN3 submitted 2026-08-06 cs.CV cs.AI

Yushe Cao , Shikun Feng , Fei Shen , Haikuo Peng , Jianqiang Xia , Yiheng Zhu , Dianxi Shi , Chun Yu This is my paper · ORCID
classification cs.CVcs.AI
keywords videovirtualtry-ondiffusiontransformermultimodallargelanguagemodelsemanticconditioningend-to-endgenerationgarmenttransferimplicitguidanceflowmatching
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

Video virtual try-on—putting a target garment onto a person in motion—is usually done by inpainting inside masks, using parse maps, pose estimates, and warped garments as crutches. UniVVT instead claims the whole task can be reframed as semantically conditioned video generation: a multimodal large language model reads the source video, the target garment, and a text instruction into compact tokens, and a diffusion video transformer follows those tokens to produce the try-on video. If this holds, the fragile geometric preprocessing that can leak source clothing or leave boundary artifacts is no longer needed at inference, and deployment simplifies to one end-to-end model. The paper reports best scores on the ViViD-S video benchmark and strong image try-on numbers on DressCode and VITON-HD.

What carries the argument

The load-bearing object is the task-aware latent $z_{\text{task}}$, produced by $N=512$ learnable task queries that attend over the MLLM-encoded source video, target garment, and instruction. A semantic bridge $P_{\theta}$ maps $z_{\text{task}}$ to generator-compatible tokens $c_{\text{proj}}$, which are injected with a frozen CLIP garment feature into every DiT block via cross-attention over the concatenated key/value sequence. This mechanism lets one model decide where to edit and how to preserve motion without ever seeing a mask or pose; the paper shows the task tokens' cross-attention concentrating on the editable garment region and staying localized across frames.

What would settle it

Compute the distribution gap between the synthetic training sources and real source videos (for example, FID in the VAE latent space), or run UniVVT on a genuine paired video set in which the same person re-films the same motion wearing a different garment; a large gap or a sharp drop in ViViD-S-style metrics would point to the synthetic-source foundation as the fragile link.

Watch

Extended reading notes

Core claim

The paper's central claim is that the what, where, and how of a virtual try-on edit can be encoded implicitly in a few hundred latent tokens, replacing explicit masks, pose maps, and warping fields. UniVVT builds a scene-task perceiver on an MLLM, prepends 512 learnable task queries to the multimodal sequence, and takes their hidden states as the task representation; a lightweight MLP semantic bridge projects these tokens into the conditioning space of a latent video diffusion transformer. During denoising, the source-video latent is concatenated with the noisy latent, the garment latent is appended, and the projected task tokens plus a frozen CLIP garment feature are injected through concatenated cross-attention. The generator is trained with conditional flow matching under a three-stage curriculum—semantic alignment with a frozen generator, joint end-to-end adaptation, and flexible-resolution refinement. The authors present state-of-the-art scores on ViViD-S and competitive image benchmarks as evidence that implicit semantic guidance is a viable replacement for geometric preprocessing.

Load-bearing premise

The load-bearing premise is that the synthetic source videos generated by inpainting real targets with alternative garments, then manually filtered, faithfully represent the real source videos UniVVT will encounter at inference; the paper reports no retention statistics, no distribution-shift measure, and no filtering-strictness ablation to support this.

Editorial extensions

If this is right

  • Drop masks, parsing, DensePose, and warping from inference: the paper's design makes the whole preprocessing stage unnecessary for deployment.
  • Latency for conditioning scales almost flat with video length; the paper reports a 17.7–38.1x speedup in conditioning overhead as videos grow from 30 to 90 frames.
  • The same architecture covers image try-on as a single-frame case, matching dedicated image methods on DressCode and VITON-HD while still doing video.
  • Three-stage progressive training stabilizes coupling between a pretrained MLLM and a pretrained video DiT, so parallel pretrained components can be steered into a new task without catastrophic forgetting.

Reading between the lines

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

  • The synthetic-source training recipe is the biggest unmeasured risk: the paper reports no retention rate from manual filtering and no distribution-shift metric, so a natural next experiment is to quantify how well the FLUX/Wan inpainting sources approximate real source videos and to re-train on real paired videos if available.
  • The same implicit-guidance mechanism could be lifted to other localized video edits—accessory changes, hairstyle changes, or scene-object replacement—since the task tokens appear to localize the edit region without explicit masks.
  • Cross-category transfer (dress to top, long to short) is acknowledged as ambiguous; adding outfit-composition instructions or a temporal memory module could disambiguate the editing extent, and the bidirectional-length results in the appendix suggest a concrete test bed.
  • The reported speedup covers preprocessing overhead, not the full diffusion sampling cost, so an end-to-end latency comparison that includes denoising would be a fairer real-world metric; the paper's own limitations note iterative denoising remains the runtime bottleneck.
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

4 major / 5 minor

Summary. This paper proposes UniVVT, a video virtual try-on framework that replaces mask-conditioned inpainting with semantic conditioning. A Qwen3-VL-based scene-task perceiver encodes the source video, target garment, and text instruction into 512 task-aware latent tokens; a lightweight MLP semantic bridge projects these tokens into the conditioning space of a Wan2.1-Fun-based diffusion transformer; and a three-stage curriculum (semantic alignment, joint adaptation, resolution refinement) trains the components. On ViViD-S the method reports the lowest VFID_I and VFID_R in paired and unpaired settings (8.3623/0.1934 and 12.3640/0.1876) and competitive image try-on results on DressCode and VITON-HD. Training triplets are constructed by synthesizing source videos with DensePose- and mask-conditioned inpainting, paired with real target garments and target videos.

Significance. If the results hold, UniVVT is a meaningful step: it removes inference-time masks, poses, and warping modules, reduces conditioning latency by 17.7–38.1×, and unifies image and video try-on in a single architecture. The paper includes useful ablations (STP, semantic bridge, LoRA, bridge capacity) and a candid limitations section. However, the main claims currently rest on an unvalidated synthetic training-source distribution and on single-run metrics with no error bars, so the significance is conditional on closing that gap.

major comments (4)
  1. [Appendix A, 'Synthesis and quality control'] The central empirical claim of the paper depends on the validity of synthetic source videos as training inputs, but the construction is validated only by manual inspection. The appendix reports no retention statistics, no quantitative measure of distribution shift between the retained synthetic sources and the real source videos used at evaluation (e.g., on the ViViD-S paired split), and no ablation varying the synthesis model or filtering strictness. Because the source is produced by DensePose- and mask-conditioned inpainting and then paired with a real target, the model could in principle learn to undo inpainting artifacts or exploit near-perfect background/pose alignment rather than a general garment-transfer capability. Please add (i) retention rates per dataset, (ii) a quantitative distribution comparison between synthetic sources and real test sources, and (iii) an ablation on filtering strictness or synthesis model, ideally complemented by training or fine-tuning on real paired triplets where available.
  2. [Section 4.4, Tables 1–3] The reported state-of-the-art claims are based on single runs with no error bars or significance tests. Some margins are very small (paired VFID_I: 8.3623 vs. 8.4030 for MagicTryOn in Table 1), and on SSIM UniVVT is not the best (0.8922 vs. 0.9011). Without confidence intervals or multiple-seed runs, 'best' is not established. Please report variance over at least three seeds or bootstrap confidence intervals for the main metrics, and state the evaluation protocol (number of videos, sampling, any preprocessing) for the reported numbers.
  3. [Section 3.1 and Appendix A] The paper's 'no geometry' framing should be scoped precisely: while inference removes masks, poses, and warping, the training data are synthesized using DensePose and mask conditioning. The manuscript is not internally inconsistent on this point (it says 'at inference'), but the claim that implicit semantic guidance is validated as an alternative to geometric preprocessing is only as strong as the demonstration that the model's localization ability is not inherited from the geometry-conditioned synthetic sources. The distribution-shift analysis in my first comment is the concrete test of this concern.
  4. [Reproducibility] No code, checkpoints, or data are released, and the paper does not report random seeds or run-to-run variance. For a purely empirical state-of-the-art claim, this makes independent verification difficult. At minimum, please release the evaluation code and trained checkpoints, and document the exact metric computation, including the KID multiplier, the number of real images used for FID, and the I3D/ResNeXt feature-extraction details.
minor comments (5)
  1. [Equation (1)] The notation ExtractQ(...) is not defined; please specify how the task-query hidden states are selected from the MLLM output sequence.
  2. [Tables 4 and 6] The 'w/o STP' ablation in Table 4 and the 'w/o z_task' ablation in Table 6 report identical numbers; please clarify whether these are the same variant described twice or two distinct ablations that happen to coincide.
  3. [Figure 2] The figure caption and diagram contain placeholder artifacts ('?????', '× ℒVAE') that should be cleaned up before publication.
  4. [Section 4.3] Please state the effective total batch size (8 GPUs × 1 per GPU = 8) and confirm that gradient accumulation is or is not used, since this affects the comparability of training schedules.
  5. [Table 5] At 512×384, Stage 3 gives VFID_I 8.3623 versus Stage 2's 8.3050; the text says 'preserving competitive quality,' but this is a slight degradation at the default resolution and should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: UniVVT's claims are empirical and its components are not defined in terms of the outputs; the synthetic-source training data concern is a validation limitation, not a circular reduction.

full rationale

I walked the paper's derivation chain, focusing on the objective (Eq. 5), the task-latent extraction (Eq. 1), the semantic bridge projection (Eq. 2), and the conditioning injection (Eq. 3). None of these definitions presuppose the benchmark outcomes or the claimed 'implicit semantic guidance' result: ztask is defined as a function of the source video, garment, and instruction, and the flow-matching loss is a standard conditional velocity objective over real target latents. The paper does not fit any parameter to the test metrics and then rename that fit as a prediction; the reported VFID, FID, KID, SSIM, and LPIPS numbers are external evaluations of a trained system. There is no self-citation chain that supplies a load-bearing 'uniqueness' or 'forbidden alternative' premise; the citations to Qwen3-VL and Wan are standard pretrained-model references, not author-specific theorems. The MLLM's hidden states are used as features, rather than the paper claiming to derive a mathematical result from them. The strongest validity concern, correctly identified in the reader's take, is Appendix A: the synthetic source videos are generated with DensePose and mask conditioning and filtered only manually, with no retention statistics, distribution-shift measurement, or filtering-strictness ablation. That is a real external-validity limitation on the claim that implicit semantic guidance is a practical alternative to geometric preprocessing, because the geometry is present in training-data construction even though it is absent at inference. However, this is a domain-shift / supervision-quality concern, not a circularity: the paper never equates its conclusion with its training-data construction, and no equation reduces the claimed result to the DensePose/mask inputs by definition. Therefore, under the hard rule that circularity must be exhibited as a specific reduction or a fitted-input-renamed-as-prediction, the appropriate finding is no circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on empirical components rather than axioms: a synthetic data assumption (the weakest), pretrained model assumptions, and standard generative-model math. The free parameters are architectural and training choices that were not fully swept.

free parameters (5)
  • Number of task queries N = 512
    Chosen by hand; determines the capacity of the task-aware latent and was not swept.
  • LoRA rank = 32
    Used for both MLLM and DiT adaptation; hand-selected, not ablated.
  • Training steps per stage = 30k
    Fixed budget for each of the three stages; no study of stage duration.
  • Base training resolution and Stage 3 resolution sampling = 512x384 base; 256p to 1024p
    Resolution curriculum chosen for memory and flexibility; not optimized.
  • Semantic Bridge MLP size = 41.96M parameters
    Compared to a Transformer variant, but no sweep over MLP width and depth.
assumptions (5)
  • standard math Conditional flow matching with the linear path in Eq. (4)-(5) yields a valid generative model for the latent video distribution.
    Invoked from Esser et al. 2024; not proven in this paper but standard.
  • domain assumption Pretrained Qwen3-VL-2B-Instruct and Wan2.1-Fun-Control provide sufficiently strong multimodal and video priors for the try-on task.
    The method inherits all capabilities and biases of these base models; no zero-shot failure analysis is given.
  • ad hoc to paper Synthetic source videos generated by FLUX.1-dev and Wan2.1-Fun with DensePose and mask conditioning are valid proxies for real source videos.
    This is the core data strategy; validity is supported only by manual inspection, not by quantitative distribution analysis.
  • ad hoc to paper The 512 task queries can encode edit location and motion correspondence through self-attention without explicit geometric supervision.
    The central architectural hypothesis; supported by internal ablations and attention maps, not by an external proof or benchmark.
  • domain assumption Public benchmarks and the compared baseline numbers are reliable and directly comparable.
    All conclusions depend on the correctness of VITON-HD, DressCode, and ViViD-S labels and on the reported baseline metrics.
invented entities (1)
  • Task-aware latent tokens z_task (512 learned query embeddings)
    purpose: Aggregate what, where, and how to edit from source video, garment, and instruction into a compact conditioning signal.
    This latent is a designed representation with no falsifiable handle outside the paper; evidence is internal ablation and attention visualization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on." pith.science (2026). https://pith.science/paper/OWNM4HN3

@misc{pith2026260805745,
  author       = {Pith},
  title        = {Pith review of: UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWNM4HN3}},
  note         = {Machine review of arXiv:2608.05745}
}
read the original abstract

Video Virtual Try-On (VVT) synthesizes a video of a person wearing a target garment while preserving identity, motion, and scene dynamics. Dominant approaches cast VVT as mask-conditioned video inpainting and rely on separate modules for human parsing, pose estimation, and garment warping. This multi-stage design complicates deployment and, more critically, allows errors in explicit geometric priors to propagate irreversibly into the generated video. We present UniVVT, a unified end-to-end framework that reframes VVT as semantically conditioned video generation, eliminating mask, pose, and warping modules at inference. At its core, a scene-task perceiver built on a Multimodal Large Language Model jointly encodes the source video, target garment, and task instruction into compact, task-aware latent tokens, implicitly capturing what to transfer and where and how to transfer it. A lightweight semantic bridge then aligns these tokens with the conditioning space of a diffusion-based video generator, enabling coherent garment transfer. To robustly couple the heterogeneous components, we devise a three-stage progressive training strategy comprising semantic alignment, joint task adaptation, and flexible-resolution refinement. Extensive experiments demonstrate that UniVVT achieves state-of-the-art performance across multiple benchmarks, validating implicit semantic guidance as a simple and effective alternative to fragile geometric preprocessing for end-to-end virtual try-on.

Figures

Figures reproduced from arXiv: 2608.05745 by the authors.

Figure 1
Figure 1. Failure cases caused by inaccurate masks. Leaked [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of UniVVT. UniVVT consists of a scene-task perceiver that comprehensively understands the try-on task [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Conditioning latency versus video length. Ex [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the ViViD-S benchmark under unpaired settings. Compared with baseline methods, UniVVT [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: In-the-wild try-on results of UniVVT on dance [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Cross-attention of ztask over denoised latents. Atten￾tion focuses on the editable garment region while remaining weak on identity and background. Garment Source Image IDM-VTON OOTDiffusion CatVTON UniVVT [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Qualitative ablation of the scene-task perceiver. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison on VITON-HD (top two [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Construction of aligned training triplets using synthetic source inputs and real-target supervision. Starting from a real [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Stage 2 training-loss curves for UniVVT with and [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Image try-on under large changes in garment [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 15
Figure 15. Figure 15: Pattern and color fidelity in video try-on. [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Complete trouser replacement in video try-on. [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Source-garment leakage suppression in video try [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]
Figure 20
Figure 20. Figure 20: Dress image try-on comparison [PITH_FULL_IMAGE:figures/full_fig_p016_20.png]
Figure 21
Figure 21. Figure 21: Cross-category failure cases with ambiguous edit [PITH_FULL_IMAGE:figures/full_fig_p017_21.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 46 canonical work pages

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume =

    Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on , author =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume =

    work page

  2. [2]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    Improving diffusion models for authentic virtual try-on in the wild , author =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    work page

  3. [3]

    International Journal of Computer Vision , volume =

    Image-based virtual try-on: A survey , author =. International Journal of Computer Vision , volume =. 2025 , publisher =

    work page 2025

  4. [4]

    arXiv preprint arXiv:2411.10499 , year =

    Fitdit: Advancing the authentic garment details for high-fidelity virtual try-on , author =. arXiv preprint arXiv:2411.10499 , year =

    arXiv

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    work page

  6. [6]

    International Journal of Computer Vision , volume =

    Tryon-adapter: Efficient fine-grained clothing identity adaptation for high-fidelity virtual try-on , author =. International Journal of Computer Vision , volume =. 2025 , publisher =

    work page 2025

  7. [7]

    IEEE Transactions on Multimedia , volume =

    PG-VTON: A novel image-based virtual try-on method via progressive inference paradigm , author =. IEEE Transactions on Multimedia , volume =. 2024 , publisher =

    work page 2024

  8. [8]

    arXiv preprint arXiv:2411.18350 , year =

    Tryoffdiff: Virtual-try-off via high-fidelity garment reconstruction using diffusion models , author =. arXiv preprint arXiv:2411.18350 , year =

    arXiv

Show all 67 references
  1. [9]

    arXiv preprint arXiv:2506.21270 , year =

    Video Virtual Try-on with Conditional Diffusion Transformer Inpainter , author =. arXiv preprint arXiv:2506.21270 , year =

    arXiv

  2. [10]

    arXiv preprint arXiv:2501.11325 , year =

    Catv2ton: Taming diffusion transformers for vision-based virtual try-on with temporal concatenation , author =. arXiv preprint arXiv:2501.11325 , year =

    arXiv

  3. [11]

    arXiv preprint arXiv:2508.02807 , year =

    Dreamvvt: Mastering realistic video virtual try-on in the wild via a stage-wise diffusion transformer framework , author =. arXiv preprint arXiv:2508.02807 , year =

    arXiv

  4. [12]

    arXiv preprint arXiv:2412.09822 , year =

    Dynamic try-on: Taming video virtual try-on with dynamic attention mechanism , author =. arXiv preprint arXiv:2412.09822 , year =

    arXiv

  5. [13]

    arXiv preprint arXiv:2505.21325 , year =

    MagicTryOn: Harnessing Diffusion Transformer for Garment-Preserving Video Virtual Try-on , author =. arXiv preprint arXiv:2505.21325 , year =

  6. [14]

    arXiv preprint arXiv:2510.07654 , year =

    Once Is Enough: Lightweight DiT-Based Video Virtual Try-On via One-Time Garment Appearance Injection , author =. arXiv preprint arXiv:2510.07654 , year =

  7. [15]

    arXiv preprint arXiv:2405.11794 , year =

    Vivid: Video virtual try-on using diffusion models , author =. arXiv preprint arXiv:2405.11794 , year =

    arXiv

  8. [16]

    arXiv preprint arXiv:2404.04833 , year =

    Shoemodel: Learning to wear on the user-specified shoes via diffusion model , author =. arXiv preprint arXiv:2404.04833 , year =

    arXiv

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Anydressing: Customizable multi-garment virtual dressing via latent diffusion models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  10. [18]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

    Street tryon: Learning in-the-wild virtual try-on from unpaired person images , author =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages =

  11. [19]

    arXiv preprint arXiv:2401.13795 , year =

    Diffuse to choose: Enriching image conditioned inpainting in latent diffusion models for virtual try-all , author =. arXiv preprint arXiv:2401.13795 , year =

    arXiv

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Viton: An image-based virtual try-on network , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  13. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Towards photo-realistic virtual try-on by adaptively generating-preserving image content , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  14. [22]

    CVPR Workshops , volume =

    Cp-vton+: Clothing shape and texture preserving image-based virtual try-on , author =. CVPR Workshops , volume =

  15. [23]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Principal warps: Thin-plate splines and the decomposition of deformations , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2002 , publisher =

    2002

  16. [24]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Clothflow: A flow-based model for clothed person generation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  17. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Parser-free virtual try-on via distilling appearance flows , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  18. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Tryondiffusion: A tale of two unets , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  19. [27]

    Proceedings of the ACM International Conference on Multimedia (ACM MM) , pages =

    Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on , author =. Proceedings of the ACM International Conference on Multimedia (ACM MM) , pages =

  20. [28]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    D4-VTON: Dynamic Semantics Disentangling for Differential Diffusion Based Virtual Try-On , author =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =

  21. [29]

    Proceedings of the ACM International Conference on Multimedia (ACM MM) , pages =

    Mv-ton: Memory-based video virtual try-on network , author =. Proceedings of the ACM International Conference on Multimedia (ACM MM) , pages =

  22. [30]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Fw-gan: Flow-navigated warping gan for video virtual try-on , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Clothformer: Taming video virtual try-on in all module , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  24. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Align your latents: High-resolution video synthesis with latent diffusion models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  25. [33]

    SIGGRAPH Asia 2024 Conference Papers , pages =

    Fashion-vdm: Video diffusion model for virtual try-on , author =. SIGGRAPH Asia 2024 Conference Papers , pages =

    2024

  26. [34]

    arXiv preprint arXiv:2511.21631 , year =

    Qwen3-VL Technical Report , author =. arXiv preprint arXiv:2511.21631 , year =

    arXiv

  27. [35]

    arXiv preprint arXiv:2409.12191 , year =

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author =. arXiv preprint arXiv:2409.12191 , year =

    arXiv

  28. [36]

    arXiv preprint arXiv:2601.03267 , year =

    Openai gpt-5 system card , author =. arXiv preprint arXiv:2601.03267 , year =

    arXiv

  29. [37]

    IEEE/CAA Journal of Automatica Sinica , volume =

    Exploring DeepSeek: A survey on advances, applications, challenges and future directions , author =. IEEE/CAA Journal of Automatica Sinica , volume =. 2025 , publisher =

    2025

  30. [38]

    arXiv preprint arXiv:2507.06261 , year =

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author =. arXiv preprint arXiv:2507.06261 , year =

    arXiv

  31. [39]

    arXiv preprint arXiv:2601.07372 , year =

    Conditional memory via scalable lookup: A new axis of sparsity for large language models , author =. arXiv preprint arXiv:2601.07372 , year =

    arXiv

  32. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Instructpix2pix: Learning to follow image editing instructions , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  33. [41]

    arXiv preprint , year =

    xMGIE: Advanced MLLM-Guided Image Editing with Progressive Feature Blending, Cross-Attention Masking, Identity Embeddings, and Gaussian Blurring , author =. arXiv preprint , year =

  34. [42]

    arXiv preprint , year =

    MAGiC: An LLM-Powered Multi-Agent Framework for Unleashing Visual Creativity , author =. arXiv preprint , year =

  35. [43]

    arXiv preprint arXiv:2501.15574 , year =

    Instruction tuning for story understanding and generation with weak supervision , author =. arXiv preprint arXiv:2501.15574 , year =

    arXiv

  36. [44]

    Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Omnialign-v: Towards enhanced alignment of mllms with human preference , author =. Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  37. [45]

    arXiv preprint arXiv:2510.08485 , year =

    Instructx: Towards unified visual editing with mllm guidance , author =. arXiv preprint arXiv:2510.08485 , year =

  38. [46]

    arXiv preprint arXiv:2503.20314 , year =

    Wan: Open and Advanced Large-Scale Video Generative Models , author =. arXiv preprint arXiv:2503.20314 , year =

    arXiv

  39. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    VITON-HD: High-Resolution Virtual Try-On via Misalignment-Aware Normalization , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2021 , month =

    2021

  40. [48]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    Dress Code: High-Resolution Multi-Category Virtual Try-On , author =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =

  41. [49]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Scalable diffusion models with transformers , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  42. [50]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  43. [51]

    GitHub repository , year =

    pytorch-fid: FID Score for PyTorch , author =. GitHub repository , year =

  44. [52]

    Demystifying

    Bi. Demystifying. International Conference on Learning Representations (ICLR) , year =

  45. [53]

    IEEE Transactions on Image Processing , volume =

    Image Quality Assessment: From Error Visibility to Structural Similarity , author =. IEEE Transactions on Image Processing , volume =

  46. [54]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  47. [55]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  48. [56]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Aggregated Residual Transformations for Deep Neural Networks , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  49. [57]

    arXiv preprint arXiv:2407.15886 , year =

    Catvton: Concatenation is all you need for virtual try-on with diffusion models , author =. arXiv preprint arXiv:2407.15886 , year =

    arXiv

  50. [58]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume =

    UniFit: Towards Universal Virtual Try-on with MLLM-Guided Semantic Alignment , author =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume =

  51. [59]

    arXiv preprint arXiv:2507.16010 , year =

    FW-VTON: Flattening-and-Warping for Person-to-Person Virtual Try-on , author =. arXiv preprint arXiv:2507.16010 , year =

    arXiv

  52. [60]

    arXiv preprint arXiv:1711.05101 , year =

    Decoupled weight decay regularization , author =. arXiv preprint arXiv:1711.05101 , year =

    arXiv

  53. [61]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Elucidating the design space of diffusion-based generative models , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  54. [62]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Reproducible scaling laws for contrastive language-image learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  55. [63]

    Proceedings of the International Conference on Machine Learning (ICML) , year =

    Scaling rectified flow transformers for high-resolution image synthesis , author =. Proceedings of the International Conference on Machine Learning (ICML) , year =

  56. [64]

    , author =

    Lora: Low-rank adaptation of large language models. , author =. International Conference on Learning Representations (ICLR) , volume =

  57. [65]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  58. [66]

    arXiv preprint arXiv:2605.10730 , year=

    Qwen-image-2.0 technical report , author=. arXiv preprint arXiv:2605.10730 , year=

    arXiv

  59. [67]

    2: Frontier Visual Intelligence , author=

    FLUX. 2: Frontier Visual Intelligence , author=. https://bfl. ai/blog/flux-2 , year=

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.