MFKP-VVTO: Mask-Free Video Virtual Try-On via Keyframe Guidance

Video virtual try-on aims to accurately transfer a target garment image onto a person in a video, generating try-on results that seamlessly align with human poses and dynamic motion. However, existing methods rely heavily on mask priors, generalize poorly, suffer from unstable temporal semantics, and lack authentic paired training data. Consequently, they yield suboptimal generation performance in challenging scenarios involving crossed limbs, large-scale pose variations, and self-occlusions. To address these limitations, we propose a keyframe-guided mask-free video virtual try-on framework, termed MFKP-VVTO. The proposed method comprises modules for pseudo-pair construction, dual-metric adaptive keyframe sampling (DA-KFS), mask-free keyframe try-on, and keyframe-guided video generation. Specifically, we leverage the pre-trained ViViD model to construct pseudo-paired samples, effectively alleviating the scarcity of real annotated data. DA-KFS integrates weighted joint distances and pose feature similarities to select representative keyframes, capturing global motion through sparse anchor frames. Furthermore, we design a mask-free try-on module built on SD-UNet to fuse dual-modal representations from CLIP and Qwen-VL and characterize garment textures and structures. We also introduce an inter-keyframe garment consistency loss that acts strictly as a regional constraint during training and does not participate in forward inference, thereby satisfying the mask-free inference requirement. Finally, the high-quality keyframe try-on results serve as temporal conditions to steer the video diffusion model, facilitating inter-frame semantic propagation. Extensive experiments on the ViViD, VVT, VITON-HD, and DressCode datasets show that MFKP-VVTO outperforms competitive baselines in occlusion and complex motion scenarios, demonstrating superior generalization and temporal coherence and providing an effective solution for mask-free video virtual try-on under complex dynamic conditions.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11417454/v1
Primary Topic
Generative Adversarial Networks and Image Synthesis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

MFKP-VVTO: Mask-Free Video Virtual Try-On via Keyframe Guidance

Shengling Geng, Weihua Pu, Xiangyan Fu, Xiaojuan Dang et al.
Research Square
Generative Adversarial Networks and Image Synthesis
preprint

MFKP-VVTO: Mask-Free Video Virtual Try-On via Keyframe Guidance

Shengling Geng, Weihua Pu, Xiangyan Fu, Xiaojuan Dang, Feizhuan Fan
preprint en

Abstract

Video virtual try-on aims to accurately transfer a target garment image onto a person in a video, generating try-on results that seamlessly align with human poses and dynamic motion. However, existing methods rely heavily on mask priors, generalize poorly, suffer from unstable temporal semantics, and lack authentic paired training data. Consequently, they yield suboptimal generation performance in challenging scenarios involving crossed limbs, large-scale pose variations, and self-occlusions. To address these limitations, we propose a keyframe-guided mask-free video virtual try-on framework, termed MFKP-VVTO. The proposed method comprises modules for pseudo-pair construction, dual-metric adaptive keyframe sampling (DA-KFS), mask-free keyframe try-on, and keyframe-guided video generation. Specifically, we leverage the pre-trained ViViD model to construct pseudo-paired samples, effectively alleviating the scarcity of real annotated data. DA-KFS integrates weighted joint distances and pose feature similarities to select representative keyframes, capturing global motion through sparse anchor frames. Furthermore, we design a mask-free try-on module built on SD-UNet to fuse dual-modal representations from CLIP and Qwen-VL and characterize garment textures and structures. We also introduce an inter-keyframe garment consistency loss that acts strictly as a regional constraint during training and does not participate in forward inference, thereby satisfying the mask-free inference requirement. Finally, the high-quality keyframe try-on results serve as temporal conditions to steer the video diffusion model, facilitating inter-frame semantic propagation. Extensive experiments on the ViViD, VVT, VITON-HD, and DressCode datasets show that MFKP-VVTO outperforms competitive baselines in occlusion and complex motion scenarios, demonstrating superior generalization and temporal coherence and providing an effective solution for mask-free video virtual try-on under complex dynamic conditions.

Research Square
Qinghai Normal University (CN), Shaanxi Fashion Engineering University (CN)
Generative Adversarial Networks and Image Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon