ACM MM Submission 4444 | Under Review

BooM-VVT: Boosting Mask-Free Video Virtual Try-On
with Image-Level Pseudo Data

A mask-free video virtual try-on framework leveraging image-level pseudo data for robust try-on region localization.

Anonymous Author(s)
Anonymous Institution
Under double-blind review
BooM-VVT teaser figure showing challenging real-world scenarios and diverse try-on tasks
Figure 1. BooM-VVT generates high-fidelity and temporally coherent virtual try-on videos without requiring masks. The left panel shows robustness in challenging real-world scenarios, while the right panel highlights diverse try-on tasks enabled by OmniView.

Abstract

Video virtual try-on (VVT) aims to generate realistic videos of a person wearing a target garment. Recent methods leverage a keyframe-driven video generation paradigm to improve in-the-wild performance, yet they still rely on masks to localize try-on regions, making them vulnerable to large motions and severe occlusions. Although mask-free image-based try-on methods have shown promising results by leveraging large-scale pseudo data, extending this paradigm to videos remains difficult, as constructing video-level pseudo data is prohibitively expensive.


To address these challenges, we propose BooM-VVT, a mask-free VVT framework built upon the keyframe-driven paradigm. BooM-VVT introduces a multi-stage training strategy that uses image-level pseudo data for mask-free localization learning, Garment-Sensitive Keyframe Sampling to select garment-relevant keyframes, Frame-Shared 3D-RoPE for accurate garment-detail transfer, and OmniView, a large-scale multi-view garment dataset for diverse try-on tasks.

Key Contributions

Multi-Stage Training

Leverages image-level pseudo data for mask-free try-on region localization, substantially reducing the need for costly video-level pseudo data.

Garment-Sensitive Keyframe Sampling

Selects keyframes based on garment-relevant body regions to better capture and preserve garment appearance across frames.

Frame-Shared 3D-RoPE

Establishes spatiotemporal correspondences between keyframes and target frames for more accurate garment-detail transfer.

OmniView Dataset

A large-scale multi-view try-on dataset with 6,110 samples supporting diverse viewpoints, categories, and try-on tasks.

Visual Comparisons

Challenging Real-World Scenarios

Large motion, object interaction, and severe occlusion.

Large Motion

Large Motion 1

Large Motion 2

Object Interaction

Object Interaction 1

Object Interaction 2

Severe Occlusion

Severe Occlusion 1

Severe Occlusion 2

Diverse Try-on Tasks

Multi-view, cross-category, and layered try-on supported by OmniView.

Multi-View Try-on

Multi-View Try-on 1

Multi-View Try-on 2

Cross-Category Try-on

Cross-Category Try-on 1

Cross-Category Try-on 2

Layered Try-on

Layered Try-on 1

Layered Try-on 2

Method

Overview of the BooM-VVT network pipeline
Figure 3. BooM-VVT consists of keyframe try-on and keyframe-guided try-on video generation. Informative keyframes are selected, converted into try-on images, and then used to guide the full video generation process.

Given an input video and target garment category, BooM-VVT first applies Garment-Sensitive Keyframe Sampling to choose frames that best expose garment-relevant body regions and complementary viewpoints. The selected keyframes are processed by a multi-frame try-on model, optionally conditioned on front- and back-view garment images, to obtain consistent keyframe try-on images. These keyframe results are then encoded as appearance tokens for the video generation model, where pose guidance, location and appearance LoRA modules, and Frame-Shared 3D-RoPE jointly help transfer garment details to the full video while preserving motion and background.

Multi-stage training strategy for BooM-VVT
Figure 4. The multi-stage strategy trains keyframe-driven video generation, learns mask-free localization from image-level pseudo data, and fine-tunes with video-level pseudo data for temporal coherence.

The training pipeline separates appearance transfer, mask-free localization, and temporal refinement. Stage 1 uses unpaired human videos with training-time garment masks to learn keyframe-driven video generation and reliable garment appearance propagation. Stage 2 replaces expensive video-level supervision with image-level pseudo pairs, forcing the model to infer the editable try-on region without explicit masks while preserving the appearance prior learned in Stage 1. Stage 3 then synthesizes and filters a smaller set of video-level pseudo samples, using them to fine-tune the mask-free model for higher-quality and temporally coherent video try-on.

BibTeX

@article{boomvvt2025,
  title   = {BooM-VVT: Boosting Mask-Free Video Virtual
             Try-On with Image-Level Pseudo Data},
  author  = {Anonymous},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2025}
}