Pith. sign in

REVIEW 4 major objections 4 minor 35 references

Dress&Dance presents a video diffusion model that turns a single user photo, a garment image, and a reference motion clip into a five-second try-on video.

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 ↗

A video diffusion framework that unifies text, image, and video conditioning through attention to produce high-resolution virtual try-on videos with reference-driven motion.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Promising video try-on system, but the same-identity training/evaluation setup and an FID mismatch undercut the headline 'outperforms' claim. the 4 major comments →

arxiv 2508.21070 v1 pith:BRCCMGA2 submitted 2025-08-28 cs.CV cs.LG

Dress&Dance: Dress up and Dance as You Like It - Technical Preview

Jun-Kun Chen , Aayush Bansal , Minh Phuoc Vo , Yu-Xiong Wang This is my paper
classification cs.CV cs.LG
keywords virtual try-onvideo diffusionmotion transfergarment transfercross-attention conditioningimage-to-video generationmulti-garment try-onvideo refiner
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 virtual try-on and motion transfer belong in one end-to-end video diffusion model rather than in a pipeline that first makes a static try-on image and then animates it. The two-stage approach loses garment information when body parts occlude clothing, and text prompts cannot convey dance-like motion precisely. Dress&Dance instead feeds the user image, target garment, optional text, and a reference video into a shared attention-based conditioning network, so the garment remains visible to the model at every generated frame. Training relies on synthetic triplets: because large-scale paired video of the same person wearing different garments is scarce, the authors create ground truth by cross-matching frames of the same person wearing different garments, then train with a staged resolution schedule. If the approach works as reported, it gives users a five-second, high-resolution video where they wear chosen clothing and move exactly like the reference, including simultaneous top-and-bottom try-on and garments borrowed from other people's photos.

Core claim

Dress&Dance's central claim is that a single video diffusion model can perform virtual try-on and motion transfer at the same time, taking a user image, one or more garment images, optional text, and a reference motion video as inputs and rendering a five-second, 24 FPS, 1152×720 video of the user wearing the target garment and moving with the reference motion. The paper argues that this joint conditioning outperforms the common two-stage recipe of image try-on followed by image animation, and that the improvement is visible both quantitatively (higher PSNR/SSIM and lower perceptual distance on captured ground-truth video, and leading scores on garment-fidelity grading with no ground truth)

What carries the argument

CondNet, a modular attention-based conditioning network. It encodes user, garment, motion-reference, and text inputs as token sequences and concatenates them with the video token sequence for cross-attention, giving every input pixel a connection to every generated pixel. This is what carries the argument: it lets garment information stay visible for the entire video, avoids first-frame error propagation, and handles flat garment images, worn garments, and multiple garments in the same conditioning pathway. Supporting it are two training mechanisms: synthetic triplet generation (cross-matching frames of the same person wearing different garments to produce unpaired ground-truth video) and a

Load-bearing premise

The load-bearing premise is that pairing the same person wearing different garments as pseudo-ground-truth teaches the model genuine garment transfer; if the model instead uses person identity to cheat the pairing, the central claim of faithful unpaired try-on fails.

What would settle it

Verify on a captured set where the user's hand covers part of the garment in the first frame and then moves away: if Dress&Dance does not reconstruct the hidden garment pattern while a variant that sees the garment at every frame does, the claimed mechanism is not doing the work. Alternatively, train the identical architecture on real paired video only, without synthetic triplets; if the quantitative and qualitative results do not change, the synthetic-supervision story is falsified.

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

If this is right

  • A two-stage 'try-on then animate' pipeline is no longer a requirement: garment, user, and motion are conditioned jointly in one diffusion pass, avoiding error propagation from the first frame.
  • The system accepts a garment as a flat image, as a worn garment on another person, or as multiple garments for simultaneous try-on, so try-on is not restricted to one clothing image.
  • Because the garment image stays visible to the model for every generated frame, patterns occluded by a hand in the first frame can reappear correctly once the hand moves — information that two-stage baselines lose.
  • Output at 1152×720, 24 FPS, 5 seconds exceeds the resolution and length of prior video try-on models, and the video refiner doubles the frame rate from 8 to 24 FPS.
  • Motion control is captured from an exemplar video, such as a dance clip, which the paper argues is more faithful than text-only motion prompts used by commercial generators.

Where Pith is reading between the lines

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

  • The same cross-attention unification could generalize beyond clothing: any video-generation task that needs several independent image conditions, such as identity, object, and scene, could reuse this token-sequence design.
  • The synthetic-triplet recipe — record each actor in multiple garment variants, then re-pair the variants to manufacture supervision — is a transferable way to create unpaired ground truth in other domains where paired video is scarce.
  • A natural test the paper does not run is stress-testing identity leakage across body shapes and skin tones: pairing a user image and a garment owner image from different demographics would reveal whether the model truly transfers the garment or copies the source person's body.
  • If the 8-to-24 FPS refiner generalizes, the same progressive training recipe could lower the cost of high-resolution video diffusion beyond virtual try-on.
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 / 4 minor

Summary. The paper proposes Dress&Dance, a video diffusion framework for virtual try-on that takes a single user image, a garment image (flat, segmented from another person, or a combination of garments), a reference motion video, and an optional text prompt, and generates a 1152×720, 24 FPS, 5-second video. The core technical contribution is CondNet, an attention-based conditioning network that unifies image, video, and text condition tokens. Training uses a multi-stage progressive scheme with a garment warm-up curriculum, and synthetic triplets are constructed from a captured video dataset to provide supervision. The paper claims state-of-the-art results against open-source and commercial baselines, supported by PSNR/SSIM/LPIPS on a private captured dataset, a GPT-based multimodal judge on Internet videos, and qualitative comparisons.

Significance. If the claims are validated, the work is significant: it addresses a practically important task—high-resolution, temporally coherent video virtual try-on with motion control—and proposes a clean architectural idea (unified attention-based conditioning) that could generalize beyond try-on. The multi-stage training strategy and the use of synthetic triplets to avoid mask/pose intermediates are also valuable. The paper ships qualitative evidence and a real captured dataset with ground-truth frames, which is more than a pure system note. However, the current evidence does not establish the headline claim of outperforming commercial systems: the main quantitative metrics are computed under a same-identity shortcut, the FID numbers in the paper contradict the claim, and the GPT-based evaluation is not validated against human judgment. The technical novelty is plausible, but the evaluation must be reworked before the central claim can be accepted.

major comments (4)
  1. [§4 Datasets] The synthetic captured-dataset triplets are built by 'cross-matching the frames of different garments of the same person' (Sec. 4). Thus every training and Table 1 evaluation triplet pairs a user image and a target garment video from the same identity. At inference the method is used with a user image of one person and a garment from another (Sec. 1, Fig. 2c), so the model can satisfy the training objective by copying identity/body cues from the garment-video frames rather than learning garment-only transfer; the same-identity correlation makes this shortcut undetectable. The PSNR/SSIM/LPIPS numbers in Table 1 therefore do not measure cross-person generalization. Please add an identity-disjoint train/test split and a quantitative identity-leakage check, or explicitly restate the claim as same-person try-on.
  2. [§4.1, Table 2] The abstract and Sec. 4.1 claim Dress&Dance 'outperforms existing open source and commercial solutions,' but Table 2's FID rows contradict this: Dress&Dance has FIDInternet 1055 and FIDCaptured 691, both worse than Kling Image Try-On + Kling Video 1.6 (982 and 655). The GPTOverall row also places full Dress&Dance (84.95) below Dress&DanceImage Try-On + Kling Video 1.6 (85.85). The claims need to be reconciled with these numbers, and the statement that Dress&Dance 'significantly outperforms all the baselines in garment fidelity' is not supported on the FID/overall metrics.
  3. [§4.1, Quantitative Comparison] The GPT-based evaluation uses rubrics designed by the authors and is never calibrated against human ratings. Re-grading each video 40 times and averaging removes sampling variance only if the judge is unbiased, which is not shown; no error bars, per-video variance, or statistical tests are reported, yet the text repeatedly uses 'significantly outperforms.' A human preference study (or at least human-GPT agreement on a subset) with confidence intervals is needed before claims of superiority on the Internet dataset can be accepted.
  4. [§4, Baselines] The VVT comparison to ViViD, WildFit, Tunnel Try-On, GPD-VVTO, ClothFormer, and Fashion-VDM is based only on 'tasks presented on their website' because no public code is available. These comparisons are uncontrolled and not quantified; they cannot support the statement that Dress&Dance 'significantly outperforms existing open source and commercial solutions.' The quantitative comparison in Table 1 uses a composed image-try-on-plus-animation pipeline, not the VVT baselines. Please add controlled quantitative comparisons for at least a subset of these methods or soften the superiority claim accordingly.
minor comments (4)
  1. [Abstract/Fig. 1] The title in Fig. 1 reads 'Dance as Y ou Like It'—the extra space in 'You' is a typo.
  2. [§4 Datasets] 'Unpaired multi-garment try-on data with ground truth' is an unusual pairing of terms; clarify what 'unpaired' means here, since the construction uses cross-matching frames of the same person.
  3. [Throughout] The name 'CondNets' appears in the contributions and architecture section, while 'CondNet' is used elsewhere; please be consistent.
  4. [Table 1] The table header contains 'Method Property Method Evaluation...' which is confusing; separate the properties from the metric columns for readability.

Circularity Check

2 steps flagged

Same-identity pseudo-ground-truth and self-authored GPT rubric create a partially self-referential evaluation loop; the architecture itself is not a derivation-level circularity.

specific steps
  1. fitted input called prediction [Section 4 (Datasets) and Section 4.1 (Quantitative Comparison, Table 1)]
    "Each model records the try-on videos for around 100 different sets of garments, so we can construct unpaired multi-garment try-on data with ground truth, by cross-matching the frames of different garments of the same person. ... we conduct experiments on tasks constructed from our captured dataset, and compare the generated video with the ground truth using PSNR, SSIM, and LPIPS."

    The 'ground truth' for both the training triplets and the Table 1 evaluation is defined by cross-matching frames of different garments on the same person, so user identity and target-garment wearer identity coincide by construction. The model is trained and quantitatively evaluated on the same same-identity distribution, and can satisfy the objective by copying identity/body cues from the target garment frames rather than by learning identity-agnostic garment transfer. The reported PSNR/SSIM/LPIPS therefore measure a same-person garment-swap task, not the cross-person try-on setting advertised in the abstract and Fig. 2(c). The quantitative 'prediction' is the fitted pseudo-ground-truth distribution renamed as an evaluation.

  2. other [Section 4.1 (Quantitative Comparison, Table 2)]
    "We leverage GPT [18]'s strong capability in vision-language reasoning to grade the generated videos ... We provide detailed instructions and rubrics to guide the GPT to grade, while re-grading each video 40 times and taking the average to avoid randomness and ensure fairness."

    The benchmark supporting the headline claim of outperforming Kling and Ray2 in the unconstrained internet setting is scored by GPT under rubrics written by the same authors who also supply the private test videos. The GPTTry-On superiority is therefore an evaluation loop in which the authors define the judge's criteria and the data; it is not an external, pre-registered benchmark. This does not reduce the method's architecture to a fitted constant, but it makes the strongest performance claim partially self-referential.

full rationale

There are no self-citations, no imported uniqueness theorems, no ansatz smuggled in via citation, and no renaming of known results. The core method (CondNet with attention-based multi-modal conditioning) is an empirical construction rather than a chain of derivations, so most of the paper is not circular in the strict sense. The circularity burden comes from the evaluation loop: the captured-dataset ground truth is generated by same-person cross-matching and then reused as the evaluation set, while the internet-dataset quality claims rest on a GPT grader with author-written rubrics. These are self-referential evaluation components rather than mathematical reductions, so the score is 4 rather than higher. The captured-dataset PSNR/SSIM numbers are still based on real videos, giving some independent anchor to the results.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 1 invented entities

The paper does not report any explicit numerical free parameters or hyperparameters; the model's millions of weights are trained, not hand-fitted. The a priori assumptions above are the load-bearing premises of the system.

axioms (4)
  • domain assumption DiT-based video diffusion models can synthesize temporally consistent high-resolution videos from multi-modal conditions.
    The whole method assumes that a diffusion transformer with attention conditioning can produce realistic 5s 24 FPS videos; no theoretical guarantee, only empirical claims.
  • ad hoc to paper Cross-attention between garment/user/motion tokens and generated video tokens suffices to preserve identity and garment fidelity.
    CondNet's central premise is asserted in Sec. 3 and only indirectly ablated; there is no decomposition showing which attention connections matter.
  • ad hoc to paper GPT-based grading with custom rubrics is a valid proxy for human judgment of try-on quality.
    The paper uses GPT re-graded 40 times as a primary quantitative tool (Table 2) without validating the rubric against human raters.
  • domain assumption Synthetic triplets created by cross-matching frames of the same person in different garments yield valid training supervision.
    In Sec. 4 the authors 'cross-match the frames of different garments of the same person' to create ground truth; if the model exploits identity cues instead of garment transfer, the central claim fails.
invented entities (1)
  • CondNet no independent evidence
    purpose: Attention-based conditioning network that unifies text, image, and video inputs into homogeneous attention sequences for the diffusion backbone.
    Listed as the key contribution but no code, weights, or standalone evaluation are provided; only end-to-end results.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Dress&Dance: Dress up and Dance as You Like It - Technical Preview." pith.science (2026). https://pith.science/paper/BRCCMGA2

@misc{pith2026250821070,
  author       = {Pith},
  title        = {Pith review of: Dress&Dance: Dress up and Dance as You Like It - Technical Preview},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BRCCMGA2}},
  note         = {Machine review of arXiv:2508.21070}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present Dress&Dance, a video diffusion framework that generates high quality 5-second-long 24 FPS virtual try-on videos at 1152x720 resolution of a user wearing desired garments while moving in accordance with a given reference video. Our approach requires a single user image and supports a range of tops, bottoms, and one-piece garments, as well as simultaneous tops and bottoms try-on in a single pass. Key to our framework is CondNet, a novel conditioning network that leverages attention to unify multi-modal inputs (text, images, and videos), thereby enhancing garment registration and motion fidelity. CondNet is trained on heterogeneous training data, combining limited video data and a larger, more readily available image dataset, in a multistage progressive manner. Dress&Dance outperforms existing open source and commercial solutions and enables a high quality and flexible try-on experience.

Figures

Figures reproduced from arXiv: 2508.21070 by Aayush Bansal, Jun-Kun Chen, Minh Phuoc Vo, Yu-Xiong Wang.

Figure 1
Figure 1. Figure 1: Given a single image of a user, a garment ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dress&Dance can be used in multiple ways: (a) Dress&Dance can generate complicated dancing motions; (b) Dress&Dance performs virtual try-on for both top and bottom garments simultaneously; and (c) Dress&Dance can input garments worn by other users [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Given a user image, a desired garment, and an indicated reference video, we extract a detailed text description via GPT [ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Our Dress&Dance supports transferring a garment from another given image via segmentation, regardless of the pose of that image. Notably, our Dress&Dance generates high-resolution videos at 1152 × 720 with clear appearance and more details, while the baseline Fashion-VDM [12] exhibits color fading and generates low-resolution 512 × 384 videos. seamlessly integrates garment try-on within video genera￾tion, … view at source ↗
Figure 5
Figure 5. Figure 5: Dress&Dance supports multi-modal conditioning through our unified CondNets architecture based on attention mechanism. In a generation, the main diffusion model generates an 8-FPS video, and then the refiner model upsamples it to 24-FPS while removing artifacts. Method Property Method Evaluation with Ground Truth on Captured Dataset Ours Commercial PSNR↑ SSIM↑ LPIPSVGG ↓ LPIPSAlexNet ↓ TPD + CogVideoX I2V 1… view at source ↗
Figure 6
Figure 6. Figure 6: Dress&Dance allows a user to dress up themselves with a desired garment, and perform the desired dance. It is difficult to express these moves by text which makes the generation of motion by Kling [1] and Ray2 [19] quite challenging. truth, by cross-matching the frames of different garments of the same person. We also collected an image dataset on the Internet with around 4M pairs of garment images paired … view at source ↗
Figure 7
Figure 7. Figure 7: Dress&Dance supports simultaneously trying on a set of top and bottom garments, while correctly understanding and represent￾ing both garments without explicit labeling. On the contrary, Kling AI [1] misrepresents the trousers as a skirt [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Dress&Dance significantly outperforms existing video virtual try-on methods, with much more detailed and precise textures and better support of transparent garments. the user image and use these models to animate accord￾ing to the text prompt. Kling AI’s platform also provides an image try-on model. We also use it to perform a try￾on for the subsequent video generation. (3) video try-on baselines. We compa… view at source ↗

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

35 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Kling ai: Next-generation ai creative studio, 2024

    Kling AI. Kling ai: Next-generation ai creative studio, 2024. 2, 3, 5, 6, 7, 8

    work page 2024

  2. [2]

    Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 2, 3, 4, 5

    work page 2023

  3. [3]

    Size does matter: Size-aware virtual try-on via clothing-oriented transformation try-on network

    Chieh-Yun Chen, Yi-Chung Chen, Hong-Han Shuai, and Wen-Huang Cheng. Size does matter: Size-aware virtual try-on via clothing-oriented transformation try-on network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7513–7522, 2023. 3

    work page 2023

  4. [4]

    Fw-gan: Flow-navigated warping gan for video virtual try-on

    Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bowen Wu, Bing-Cheng Chen, and Jian Yin. Fw-gan: Flow-navigated warping gan for video virtual try-on. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 4

    work page 2019

  5. [5]

    Videoshop: Localized semantic video editing with noise-extrapolated diffusion inversion, 2024

    Xiang Fan, Anand Bhattad, and Ranjay Krishna. Videoshop: Localized semantic video editing with noise-extrapolated diffusion inversion, 2024. 3 7 Method Property Method Evaluation of Try-On and Visual Quality Ours Commercial GPTTry-On↑ GPTUser↑ GPTMotion↑ GPTVisual↑ GPTOverall↑ FIDInternet↓ FIDCaptured↓ TPD + CogVideoX I2V 69.67 73.98 68.15 65.45 68.64 114...

    work page 2024

  6. [6]

    Vivid: Video virtual try-on using diffusion models,

    Zixun Fang, Wei Zhai, Aimin Su, Hongliang Song, Kai Zhu, Mao Wang, Yu Chen, Zhiheng Liu, Yang Cao, and Zheng- Jun Zha. Vivid: Video virtual try-on using diffusion models,

  7. [7]

    Parser-free virtual try-on via distilling appearance flows

    Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8485–8493, 2021. 3

    work page 2021

  8. [8]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 4

    work page 2014

  9. [9]

    Reuse and diffuse: Iterative denoising for text-to-video generation, 2023

    Jiaxi Gu, Shicong Wang, Haoyu Zhao, Tianyi Lu, Xing Zhang, Zuxuan Wu, Songcen Xu, Wei Zhang, Yu-Gang Jiang, and Hang Xu. Reuse and diffuse: Iterative denoising for text-to-video generation, 2023. 2

    work page 2023

  10. [10]

    Wildvidfit: Video virtual try- on in the wild via image-based controlled diffusion models

    Zijian He, Peixin Chen, Guangrun Wang, Guanbin Li, Philip HS Torr, and Liang Lin. Wildvidfit: Video virtual try- on in the wild via image-based controlled diffusion models. arXiv preprint arXiv:2407.10625, 2024. 4, 6

    Pith/arXiv arXiv 2024

  11. [11]

    Cloth- former: Taming video virtual try-on in all module

    Jianbin Jiang, Tan Wang, He Yan, and Junhui Liu. Cloth- former: Taming video virtual try-on in all module. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 6

    work page 2022

  12. [12]

    Fashion-vdm: Video diffusion model for virtual try-on

    Johanna Karras, Yingwei Li, Nan Liu, Luyang Zhu, Innfarn Yoo, Andreas Lugmayr, Chris Lee, and Ira Kemelmacher- Shlizerman. Fashion-vdm: Video diffusion model for virtual try-on. In Proceedings of ACM SIGGRAPH Asia 2024, De- cember 2024. 4, 6, 7

    work page 2024

  13. [13]

    Stableviton: Learning semantic corre- spondence with latent diffusion model for virtual try-on

    Jeongho Kim, Gyojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic corre- spondence with latent diffusion model for virtual try-on. In CVPR, 2024. 3

    work page 2024

  14. [14]

    Shineon: Illuminating design choices for practical video-based virtual clothing try-on, 2020

    Gaurav Kuppa, Andrew Jong, Vera Liu, Ziwei Liu, and Teng Moh. Shineon: Illuminating design choices for practical video-based virtual clothing try-on, 2020. 4

    work page 2020

  15. [15]

    High-resolution virtual try-on with misalignment and occlusion-handled conditions

    Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Eu- ropean Conference on Computer Vision , pages 204–219. Springer, 2022. 2, 3, 5

    work page 2022

  16. [16]

    Evaluating text-to-visual generation with image-to-text gen- eration

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text gen- eration. In ECCV, 2024. 7

    work page 2024

  17. [17]

    SDEdit: Guided image synthesis and editing with stochastic differential equa- tions

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equa- tions. In ICLR, 2022. 3

    work page 2022

  18. [18]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7

    Pith/arXiv arXiv 2023

  19. [19]

    Ray 2 - advanced ai video generation tool, 2025

    Ray2. Ray 2 - advanced ai video generation tool, 2025. 2, 3, 5, 6, 7, 8

    work page 2025

  20. [20]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 2

    2022

  21. [21]

    Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2024

    Fengyuan Shi, Jiaxi Gu, Hang Xu, Songcen Xu, Wei Zhang, and Limin Wang. Bivdiff: A training-free framework for general-purpose video synthesis via bridging image and video diffusion models, 2024. 3

    work page 2024

  22. [22]

    Stablegar- ment: Garment-centric generation via stable diffusion

    Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegar- ment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783, 2024. 3

    Pith/arXiv arXiv 2024

  23. [23]

    Gpd-vvto: Preserving garment details in video virtual try-on

    Yuanbin Wang, Weilun Dai, Long Chan, Huanyu Zhou, Aixi Zhang, and Si Liu. Gpd-vvto: Preserving garment details in video virtual try-on. In Proceedings of the 32nd ACM International Conference on Multimedia , MM ’24, page 7133–7142, New York, NY , USA, 2024. Association for Computing Machinery. 4, 6

    work page 2024

  24. [24]

    Gp- vton: Towards general purpose virtual try-on via collabora- tive local-flow global-parsing learning

    Zhenyu Xiel, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp- vton: Towards general purpose virtual try-on via collabora- tive local-flow global-parsing learning. pages 23550–23559, 06 2023. 2, 3, 5

    work page 2023

  25. [25]

    Tryon-adapter: Efficient fine-grained clothing identity adaptation for high- fidelity virtual try-on

    Jiazheng Xing, Chao Xu, Yijie Qian, Yang Liu, Guang Dai, Baigui Sun, Yong Liu, and Jingdong Wang. Tryon-adapter: Efficient fine-grained clothing identity adaptation for high- fidelity virtual try-on. International Journal of Computer Vi- sion, pages 1–22, 2025

    work page 2025

  26. [26]

    Oot- diffusion: Outfitting fusion based latent diffusion for control- lable virtual try-on

    Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot- diffusion: Outfitting fusion based latent diffusion for control- lable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 2, 3, 5

    Pith/arXiv arXiv 2024

  27. [27]

    Tunnel try-on: Excavating spatial-temporal tunnels for high-quality virtual try-on in videos, 2024

    Zhengze Xu, Mengting Chen, Zhao Wang, Linyu Xing, Zhonghua Zhai, Nong Sang, Jinsong Lan, Shuai Xiao, and Changxin Gao. Tunnel try-on: Excavating spatial-temporal tunnels for high-quality virtual try-on in videos, 2024. 4, 6 8

    work page 2024

  28. [28]

    Texture-preserving diffusion models for high-fidelity virtual try-on, 2024

    Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on, 2024. 2, 3, 5

    work page 2024

  29. [29]

    D 4-vton: Dynamic semantics disentangling for differential diffusion based virtual try-on

    Zhaotong Yang, Zicheng Jiang, Xinzhe Li, Huiyu Zhou, Junyu Dong, Huaidong Zhang, and Yong Du. D 4-vton: Dynamic semantics disentangling for differential diffusion based virtual try-on. In European Conference on Computer Vision, pages 36–52. Springer, 2024. 3

    work page 2024

  30. [30]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5

    Pith/arXiv arXiv 2024

  31. [31]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 586–595, 2018. 7

    work page 2018

  32. [32]

    I2vgen-xl: High-quality image-to-video syn- thesis via cascaded diffusion models

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video syn- thesis via cascaded diffusion models. 2023. 2, 3, 5

    work page 2023

  33. [33]

    MV-TON: Memory-based Video Virtual Try-on network

    Xiaojing Zhong, Zhonghua Wu, Taizhe Tan, Guosheng Lin, and Qingyao Wu. MV-TON: memory-based video virtual try-on network. CoRR, abs/2108.07502, 2021. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  34. [34]

    M&m vto: Multi- garment virtual try-on and editing

    Luyang Zhu, Yingwei Li, Nan Liu, Hao Peng, Dawei Yang, and Ira Kemelmacher-Shlizerman. M&m vto: Multi- garment virtual try-on and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1346–1356, 2024. 3

    work page 2024

  35. [35]

    Tryondiffusion: A tale of two unets

    Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Saharia, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 4606–4615,

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.