arxiv: 2604.04934 · v2 · submitted 2026-04-06 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

Vanast: Virtual Try-On with Human Image Animation via Synthetic Triplet Supervision

Byungjun Kim, Hanbyul Joo, Hyunsoo Cha, Wonjung Woo

Authors on Pith no claims yet

Pith reviewed 2026-05-10 20:30 UTC · model grok-4.3

classification 💻 cs.CV

keywords virtual try-onhuman image animationgarment transfervideo diffusionsynthetic triplet supervisionpose guidanceidentity preservationunified synthesis

0 comments

The pith

Vanast performs virtual try-on and human animation together in one unified step using synthetic triplet data to avoid identity drift and garment distortion.

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

The paper introduces Vanast as a framework that takes a single human image, garment images, and a pose guidance video to output garment-transferred animation videos. Conventional methods split virtual try-on and pose animation into separate stages, which commonly produces identity changes, warped clothing, and front-back mismatches. Vanast instead handles the full process at once by training on large-scale synthetic triplets that include identity-preserving outfits and complete upper-plus-lower garment sets drawn from both catalog and in-the-wild sources. A Dual Module design inside video diffusion transformers keeps training stable, retains generative quality, and boosts accuracy on garments, poses, and identity while allowing zero-shot garment swaps. The result is claimed to deliver coherent, high-fidelity animations over a wide range of clothing types.

Core claim

Vanast generates garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video by executing the entire process in one unified step rather than sequential try-on and animation stages, trained on synthetic triplets that supply identity-preserving alternative outfits and full upper-lower garment combinations, and realized through a Dual Module architecture in video diffusion transformers that stabilizes training, preserves pretrained quality, and improves garment accuracy, pose adherence, and identity preservation.

What carries the argument

Synthetic triplet supervision paired with a Dual Module architecture inside video diffusion transformers, where the dual modules stabilize training and improve garment accuracy, pose adherence, and identity preservation while supporting zero-shot interpolation.

If this is right

Produces coherent videos without identity drift, garment distortion, or front-back inconsistency
Handles both upper and lower garments in a single pass
Supports zero-shot interpolation between different garments
Maintains high generative quality from the pretrained diffusion model
Works on diverse in-the-wild triplets without needing catalog images

Where Pith is reading between the lines

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

A unified single-step model could reduce the number of separate processing stages needed for online clothing visualization tools
The approach might generalize to longer or multi-view videos if the triplet pipeline is extended to include temporal consistency checks
Further work could test whether the same synthetic data strategy applies to full-body or multi-person animation settings

Load-bearing premise

The synthetic triplet data generation pipeline produces realistic and diverse training examples that transfer to real-world garment images and videos without introducing artifacts or biases that harm identity preservation or garment accuracy.

What would settle it

Side-by-side visual comparison of Vanast outputs against real captured videos of the same person in the target garments, scored for identity match and garment fidelity, would show whether artifacts or drift appear in practice.

Figures

Figures reproduced from arXiv: 2604.04934 by Byungjun Kim, Hanbyul Joo, Hyunsoo Cha, Wonjung Woo.

**Figure 1.** Figure 1: Vanast. Given a human image and one or more garment images, our method generates virtual try-on with human image animation conditioned on a pose video while preserving identity. Abstract We present Vanast, a unified framework that generates garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video. Conventional two-stage pipelines treat image-b… view at source ↗

**Figure 2.** Figure 2: Overview of Vanast Pipeline. Our Vanast framework generates virtual try-on human animation videos from a human image, garment images, and a pose video. By incorporating scalable human-image and garment-image generation pipelines, our method avoids dataset-specific constraints and trains effectively at scale. The Dual Modules architecture ensures that the three conditioning signals, human image I G′ , garme… view at source ↗

**Figure 3.** Figure 3: Samples of Synthetic Triplet Datasets. We show samples of the datasets used for generation and training. The triplet construction contributes to enabling the model to preserve identity while accurately transferring garments and producing animation videos that follow the target pose. consists of three stages: (1) selecting suitable candidate frames I G, (2) constructing proper and effective inpainting mask… view at source ↗

**Figure 4.** Figure 4: Qualitative Comparisons (Subject-to-Image-based). We compare our results with baselines constructed by combining subject-to-image generation and animation models. Our method produces the most accurate pose following and garment transfer while preserving identity with high fidelity. Internet ViViD OOTDiffusion Stab-Ani DisPose CatVTON Stab-Ani DisPose OmniTry Stab-Ani DisPose Ours GT Any2AnyTryon Stab-Ani D… view at source ↗

**Figure 5.** Figure 5: Qualitative Comparisons (Virtual Try-On-based). We compare our results with baselines formed by combining image virtual try-on models with animation models. Our method achieves the most accurate pose following and garment transfer while preserving identity with the highest fidelity. mask-conditioned dual-Unet diffusion model for referenceguided garment transfer; CatVTON [5], a diffusion-based try-on mode… view at source ↗

**Figure 6.** Figure 6: Result of Single Garment Transfer. We present virtual try-on with human image animation results generated from a single garment image. the model using LoRA layers, applied to every DiT block. This setup enables faster convergence while preserving the generative capabilities of T2V. “w/o SynthHuman” shares the same architecture as our full model but is trained without using I G′ , relying solely on I G. As … view at source ↗

**Figure 8.** Figure 8: Result of Multiple Garment Transfer. We present zero-shot garment transfer results where both upper and lower garments are transferred simultaneously. The logos and fine details of the garments are well preserved and accurately reflected in the generated animation videos. Human In-the-wild Garment Virtual Try-On and Animation [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

**Figure 7.** Figure 7: Ablation Study. We present the ablation study results for the lower garment transfer. The red box in the “Single Module” result demonstrates vulnerability to pose conditions. Both “Backbone-LoRA” and “w/o SynthHuman” fail to achieve accurate garment transfer, as indicated in blue box. In contrast, our full model produces results most similar to the ground truth. rately control pose conditions, while the “… view at source ↗

**Figure 10.** Figure 10: Result of Garment Interpolation. Without requiring any additional finetuning, our Vanast model performs zero-shot transfer of interpolated garments by GTM. γ denotes a scalar interpolation weight. 5. Conclusion We introduce Vanast, a unified framework that synthesizes garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video. By constructing … view at source ↗

read the original abstract

We present Vanast, a unified framework that generates garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video. Conventional two-stage pipelines treat image-based virtual try-on and pose-driven animation as separate processes, which often results in identity drift, garment distortion, and front-back inconsistency. Our model addresses these issues by performing the entire process in a single unified step to achieve coherent synthesis. To enable this setting, we construct large-scale triplet supervision. Our data generation pipeline includes generating identity-preserving human images in alternative outfits that differ from garment catalog images, capturing full upper and lower garment triplets to overcome the single-garment-posed video pair limitation, and assembling diverse in-the-wild triplets without requiring garment catalog images. We further introduce a Dual Module architecture for video diffusion transformers to stabilize training, preserve pretrained generative quality, and improve garment accuracy, pose adherence, and identity preservation while supporting zero-shot garment interpolation. Together, these contributions allow Vanast to produce high-fidelity, identity-consistent animation across a wide range of garment types.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Vanast unifies virtual try-on and pose animation in one diffusion step via synthetic triplets and a dual module, but the whole thing stands or falls on whether that data pipeline actually transfers to real images without drift.

read the letter

The paper's core move is to stop treating try-on and animation as two separate stages. Instead it trains a single video diffusion model on triplets that include the person, the target garments, and a pose sequence. The data pipeline generates identity-preserving outfit swaps, full upper-plus-lower garment sets, and in-the-wild combinations that do not rely on catalog photos. A Dual Module is added to the transformer to keep the base model's quality while tightening garment and pose control. That is the actual novelty: the specific triplet construction and the architectural tweak to support zero-shot garment swaps inside the same forward pass. It directly targets the identity drift and front-back problems that two-stage pipelines usually inherit from error accumulation. The approach is straightforward and the motivation is clear from the problems it lists. The main weakness is that everything rests on the synthetic triplets being close enough to real garment photos and videos. Lighting, fabric drape, body shape variation, and temporal consistency all have to match or the model will learn the wrong distribution. The abstract gives no domain-gap numbers, no real-versus-synthetic ablation, and no quantitative comparison to prior two-stage methods on standard metrics. Without those, it is hard to know whether the unified step actually delivers the claimed coherence or just moves the failure modes around. This is useful reading for anyone working on generative fashion video or diffusion-based animation. The ideas are concrete enough that a referee could check the data pipeline and run the necessary controls. I would send it to review and ask specifically for the missing ablations and real-data results.

Referee Report

1 major / 3 minor

Summary. The manuscript presents Vanast, a unified single-step framework for garment-transferred human animation video generation. Given a single reference human image, target garment images, and a pose guidance video, the method produces coherent output videos via a video diffusion transformer trained on large-scale synthetic triplets. The triplets are constructed through identity-preserving outfit variation, full upper/lower garment coverage, and in-the-wild assembly without catalog images. A Dual Module architecture is introduced to stabilize training, retain pretrained generative quality, and improve garment accuracy, pose adherence, and identity preservation while enabling zero-shot garment interpolation.

Significance. If the central claims hold, the work offers a meaningful advance over conventional two-stage virtual try-on plus animation pipelines by reducing identity drift, garment distortion, and front-back inconsistency through unified training. The synthetic triplet pipeline and Dual Module design constitute concrete technical contributions that could be adopted in related diffusion-based video synthesis tasks, provided the domain gap between synthetic training data and real inputs is shown to be manageable.

major comments (1)

[§3] §3 (Synthetic Triplet Data Generation): The unified single-step claim and the headline performance on identity-consistent animation rest on the assertion that the synthetic pipeline produces training examples whose distribution is sufficiently close to real garment images and in-the-wild pose videos. No quantitative domain-gap measurements (e.g., FID, LPIPS, or perceptual user studies between synthetic and real triplets), no ablation of real-vs-synthetic training, and no error analysis of failure modes induced by lighting, fabric deformation, or body-shape mismatches are reported. This is load-bearing because any systematic mismatch would directly undermine the coherence advantages claimed over two-stage baselines.

minor comments (3)

[§4.2] §4.2 (Dual Module): The interaction between the two modules and the precise conditioning mechanisms for garment, pose, and identity are described at a high level; explicit equations for the combined forward process or the zero-shot interpolation procedure would improve reproducibility.
[Figure 3] Figure 3 and Table 1: Captions and axis labels should explicitly state whether metrics are computed on synthetic validation triplets or on held-out real videos to allow direct assessment of the transfer claim.
[§2] §2 (Related Work): The discussion of prior virtual try-on and video diffusion methods is adequate but could more clearly delineate the precise novelty of the full-upper/lower triplet construction relative to existing paired try-on datasets.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential of our unified framework and technical contributions. We address the major comment on the synthetic triplet pipeline below and will strengthen the manuscript accordingly.

read point-by-point responses

Referee: [§3] §3 (Synthetic Triplet Data Generation): The unified single-step claim and the headline performance on identity-consistent animation rest on the assertion that the synthetic pipeline produces training examples whose distribution is sufficiently close to real garment images and in-the-wild pose videos. No quantitative domain-gap measurements (e.g., FID, LPIPS, or perceptual user studies between synthetic and real triplets), no ablation of real-vs-synthetic training, and no error analysis of failure modes induced by lighting, fabric deformation, or body-shape mismatches are reported. This is load-bearing because any systematic mismatch would directly undermine the coherence advantages claimed over two-stage baselines.

Authors: We agree that quantifying the domain gap is essential to substantiate the benefits of unified single-step training. The synthetic triplet pipeline was explicitly engineered to narrow this gap through identity-preserving outfit variation, full upper/lower garment coverage, and in-the-wild assembly from diverse sources. However, the initial submission indeed omitted the requested quantitative metrics, ablations, and error analysis. In the revised manuscript we will add: (1) FID and LPIPS comparisons between our synthetic triplets and real garment-pose sequences; (2) an ablation study reporting performance when training on synthetic data alone versus any obtainable real or mixed triplets; and (3) a dedicated error-analysis subsection with qualitative examples of failure modes arising from lighting mismatches, fabric deformation, and body-shape variations. These additions will directly support the coherence and identity-preservation claims relative to two-stage baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the Vanast derivation chain

full rationale

The paper introduces a unified single-step framework for garment-transferred human animation, enabled by an explicitly constructed synthetic triplet data generation pipeline (identity-preserving alternative outfits, full upper/lower garment triplets, in-the-wild assembly) and a new Dual Module video diffusion transformer architecture. These elements are presented as novel contributions rather than derivations that reduce by construction to fitted parameters, self-definitions, or self-cited uniqueness theorems. No equations or load-bearing self-citations appear in the abstract or described claims; the method trains on generated data in a standard non-circular ML manner, with performance claims resting on external experimental validation rather than internal renaming or forced equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical derivations or physical axioms present. The work rests on standard assumptions of diffusion model training and the effectiveness of synthetic data for supervision.

pith-pipeline@v0.9.0 · 5495 in / 1219 out tokens · 32875 ms · 2026-05-10T20:30:30.980478+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
We present Vanast, a unified framework that generates garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video... Dual Module architecture for video diffusion transformers
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
Our data generation pipeline includes generating identity-preserving human images in alternative outfits... Dual Module architecture... zero-shot garment interpolation

Reference graph

Works this paper leans on

40 extracted references · 17 canonical work pages · 6 internal anchors

[1]

S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 4, 5

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

H. Cha, B. Kim, and H. Joo. Pegasus: Personalized generative 3d avatars with composable attributes. InCVPR, 2024. 2

2024
[3]

H. Cha, B. Kim, and H. Joo. Durian: Dual reference image-guided portrait animation with attribute transfer.arXiv preprint arXiv:2509.04434, 2025. 2

work page internal anchor Pith review Pith/arXiv arXiv 2025
[4]

H. Cha, I. Lee, and H. Joo. Perse: Personalized 3d generative avatars from a single portrait. InCVPR, 2025. 4, 5

2025
[5]

Zheng Chong, Yanwei Lei, Shiyue Zhang, Zhuandi He, Zhen Wang, Xujie Zhang, Xiao Dong, Yiling Wu, Dongmei Jiang, and Xiaodan Liang

Z. Chong, X. Dong, H. Li, S. Zhang, W. Zhang, X. Zhang, H. Zhao, D. Jiang, and X. Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models.arXiv preprint arXiv:2407.15886, 2024. 1, 2, 6

work page arXiv 2024
[6]

H. Dong, X. Liang, X. Shen, B. Wu, B.-C. Chen, and J. Yin. Fw-gan: Flow-navigated warping gan for video virtual try-on. InICCV, 2019. 2

2019
[7]

Z. Fang, W. Zhai, A. Su, H. Song, K. Zhu, M. Wang, Y . Chen, Z. Liu, Y . Cao, and Z.-J. Zha. Vivid: Video virtual try-on using diffusion models.arXiv preprint arXiv:2405.11794,

work page arXiv
[8]

Y . Feng, L. Zhang, H. Cao, Y . Chen, X. Feng, J. Cao, Y . Wu, and B. Wang. Omnitry: Virtual try-on anything without masks.arXiv preprint arXiv:2508.13632, 2025. 1, 2, 6

work page arXiv 2025
[9]

H. Guo, B. Zeng, Y . Song, W. Zhang, C. Zhang, and J. Liu. Any2anytryon: Leveraging adaptive position em- beddings for versatile virtual clothing tasks.arXiv preprint arXiv:2501.15891, 2025. 1, 2, 6

work page arXiv 2025
[10]

Y . Guo, C. Yang, A. Rao, Z. Liang, Y . Wang, Y . Qiao, M. Agrawala, D. Lin, and B. Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023. 3

work page internal anchor Pith review arXiv 2023
[11]

LTX-Video: Realtime Video Latent Diffusion

Y . HaCohen, N. Chiprut, B. Brazowski, D. Shalem, D. Moshe, E. Richardson, E. Levin, G. Shiran, N. Zabari, O. Gordon, et al. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024. 3

work page internal anchor Pith review arXiv 2024
[12]

Heusel, H

M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.NeurIPS, 2017. 6

2017
[13]

L. Hu. Animate anyone: Consistent and controllable image- to-video synthesis for character animation. InCVPR, 2024. 1, 3

2024
[14]

Jiang, C

Z. Jiang, C. Mao, Z. Huang, A. Ma, Y . Lv, Y . Shen, D. Zhao, and J. Zhou. Res-tuning: A flexible and efficient tuning paradigm via unbinding tuner from backbone.NeurIPS, 2023. 5

2023
[15]

Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025

Z. Jiang, Z. Han, C. Mao, J. Zhang, Y . Pan, and Y . Liu. Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025. 2, 3, 5, 6

work page arXiv 2025
[16]

J. Kim, G. Gu, M. Park, S. Park, and J. Choo. Stablevi- ton: Learning semantic correspondence with latent diffusion model for virtual try-on. InCVPR, 2024. 1, 2

2024
[17]

B. F. Labs. Flux. https://github.com/black- forest-labs/flux, 2024. 4, 5, 6

2024
[18]

H. Li, Y . Li, Y . Yang, J. Cao, Z. Zhu, X. Cheng, and L. Chen. Dispose: Disentangling pose guidance for controllable human image animation.arXiv preprint arXiv:2412.09349, 2024. 1, 3, 6

work page arXiv 2024
[19]

Z.-Y . Li, R. Du, J. Yan, L. Zhuo, Z. Li, P. Gao, Z. Ma, and M.-M. Cheng. Visualcloze: A universal image generation framework via visual in-context learning.arXiv preprint arXiv:2504.07960, 2025. 3, 6

work page arXiv 2025
[20]

L. Liu, T. Ma, B. Li, Z. Chen, J. Liu, G. Li, S. Zhou, Q. He, and X. Wu. Phantom: Subject-consistent video generation via cross-modal alignment.arXiv preprint arXiv:2502.11079,

work page arXiv
[21]

Nguyen, Q

H. Nguyen, Q. Q.-V . Nguyen, K. Nguyen, and R. Nguyen. Swifttry: Fast and consistent video virtual try-on with diffu- sion models. InAAAI, 2025. 2, 8

2025
[22]

Chatgpt (gpt-5)

OpenAI. Chatgpt (gpt-5). https://chat.openai.com,
[23]

SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M¨uller, J. Penna, and R. Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023. 4

work page internal anchor Pith review Pith/arXiv arXiv 2023
[24]

Redmon, S

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. InCVPR,
[25]

Rombach, A

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022. 3

2022
[26]

D. She, S. Fu, M. Liu, Q. Jin, H. Wang, M. Liu, and J. Jiang. Mosaic: Multi-subject personalized generation via correspondence-aware alignment and disentanglement.arXiv preprint arXiv:2509.01977, 2025. 3, 6

work page arXiv 2025
[27]

S. Tu, Z. Xing, X. Han, Z.-Q. Cheng, Q. Dai, C. Luo, and Z. Wu. Stableanimator: High-quality identity-preserving human image animation. InCVPR, 2025. 1, 3, 6

2025
[28]

T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 2, 3, 5

work page internal anchor Pith review Pith/arXiv arXiv 2025
[29]

X. Wang, S. Zhang, C. Gao, J. Wang, X. Zhou, Y . Zhang, L. Yan, and N. Sang. Unianimate: Taming unified video diffu- sion models for consistent human image animation.Science China Information Sciences, 2025. 3

2025
[30]

Z. Wang, Y . Li, Y . Zeng, Y . Fang, Y . Guo, W. Liu, J. Tan, K. Chen, T. Xue, B. Dai, et al. Humanvid: Demystifying training data for camera-controllable human image animation. NeurIPS, 2024. 5, 6

2024
[31]

S. Wu, M. Huang, W. Wu, Y . Cheng, F. Ding, and Q. He. Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160,

work page arXiv
[32]

E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers.NeurIPS, 2021. 5

2021
[33]

Y . Xu, T. Gu, W. Chen, and A. Chen. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. InAAAI, 2025. 2, 6

2025
[34]

Z. Xu, Z. Huang, J. Cao, Y . Zhang, X. Cun, Q. Shuai, Y . Wang, L. Bao, J. Li, and F. Tang. Anchorcrafter: Animate cyber- anchors saling your products via human-object interacting video generation.arXiv preprint arXiv:2411.17383, 2024. 3

work page arXiv 2024
[35]

H. Yang, R. Zhang, X. Guo, W. Liu, W. Zuo, and P. Luo. To- wards photo-realistic virtual try-on by adaptively generating- preserving image content. InCVPR, 2020. 2

2020
[36]

S. Yang, H. Li, J. Wu, M. Jing, L. Li, R. Ji, J. Liang, H. Fan, and J. Wang. Megactor-σ: Unlocking flexible mixed-modal control in portrait animation with diffusion transformer, 2025. 3

2025
[37]

Z. Yang, A. Zeng, C. Yuan, and Y . Li. Effective whole-body pose estimation with two-stages distillation. InICCV, 2023. 4

2023
[38]

S. Yuan, J. Huang, X. He, Y . Ge, Y . Shi, L. Chen, J. Luo, and L. Yuan. Identity-preserving text-to-video generation by frequency decomposition. InCVPR, 2025. 3

2025
[39]

Zhang, Y

K. Zhang, Y . Zhou, X. Xu, B. Dai, and X. Pan. Diffmor- pher: Unleashing the capability of diffusion models for image morphing. InCVPR, 2024. 5

2024
[40]

S. Zhu, J. L. Chen, Z. Dai, Y . Xu, X. Cao, Y . Yao, H. Zhu, and S. Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. InECCV, 2024. 1, 3 10

2024