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

• Existing VVT methods struggle with garment fidelity and temporal consistency because paired garment–video data are scarce and scene diversity is limited.Source: §1 · p. 2
• Mask-based localization is vulnerable to large motions and occlusions, while extending image-level pseudo-data methods to video is costly because video-level pseudo data are expensive to construct.Source: §1 · p. 2
• BooM-VVT uses multi-stage training to learn mask-free try-on-region localization from image-level pseudo data, reducing reliance on video-level pseudo data.Source: §1 · p. 2
• Garment-Sensitive Keyframe Sampling selects frames using garment-relevant body regions, and Frame-Shared 3D-RoPE establishes correspondences for accurate garment appearance transfer.Source: §1 · p. 2
• OmniView provides large-scale multi-view try-on data to support reliable generation across diverse viewpoints and try-on tasks.Source: §1 · p. 2

• Existing multi-view try-on datasets are limited in scale, garment diversity, and completeness.Source: §2 · pp. 2–3
• OmniView contains 6,110 samples, each with person images from at least two viewpoints, and 88% include back-view images.Source: §2 · p. 3
• Each OmniView sample provides front- and back-view flat garment images for more complete garment references.Source: §2 · p. 3
• OmniView covers upper-body, lower-body, full-body, and an additional outerwear garment category.Source: §2 · p. 3
• The dataset includes synthesized pseudo-labeled data supporting cross-category and layered try-on.Source: §2 · p. 3

Garment-Sensitive Keyframe Sampling

• GSKS evaluates garment-relevant body regions and selects a high-capacity primary keyframe before adding frames with complementary viewpoint coverage.Source: §3.2 · p. 3, Multi-Frame Try-on Model · p. 4

Overview

• The framework selects informative keyframes, generates their try-on images, and synthesizes the final try-on video under keyframe guidance.Source: §3.1 · p. 3, Figure 3 · p. 4

Garment-Sensitive Keyframe Sampling

• For upper-body garments, GSKS considers shoulders and arms as relevant limbs and the upper torso as the relevant body region.Source: Multi-Frame Try-on Model · p. 4

Keyframe Try-on

• Selected keyframes are horizontally concatenated and processed in one image-model forward pass, enabling cross-frame information interaction before being split into individual try-on frames.Source: Multi-Frame Try-on Model · p. 4

Keyframe-Guided Video Generation

• The video generator encodes video, pose, and keyframe tokens, concatenates keyframe and video tokens, and decodes the denoised representation into the final try-on video.Source: §3.3 · p. 5

Frame-Shared 3D-RoPE

• Frame-Shared 3D-RoPE assigns aligned keyframe and video-frame tokens identical positional encodings to reduce their relative positional distance and strengthen attention.Source: §3.3 · p. 5, §3.3.1 · p. 5

Multi-Stage Training

• Multi-stage training first learns video generation with mask-provided regions and then learns mask-free localization using image-level pseudo pairs.Source: §3.3.2 · p. 5, §3.3.1 · p. 5

Evaluation Setup

• The evaluation uses ViViD-S with 180 samples and WildVVT with 100 samples covering diverse motions, scenes, and garment categories.Generated videos are assessed with FVD, SSIM, LPIPS, and judge-based quality dimensions.Source: §4.1 · pp. 6–7

Qualitative Comparison

• BooM-VVT produces more realistic and temporally coherent results than ViViD, CatV2TON, and MagicTryOn on ViViD-S and WildVVT.Competing mask-based methods show artifacts under garment-shape changes, turning motions, and severe occlusions.Source: §4.2 · p. 7

Quantitative Comparison

• Keyframe-driven BooM-VVT variants consistently outperform non-keyframe-driven baselines on both ViViD-S and WildVVT.Because public implementations of existing keyframe-driven methods are unavailable, BooM-VVT Stage 1 serves as the comparison baseline.Source: §4.3 · p. 7
• BooM-VVT achieves the best performance across all quality metrics and substantially outperforms MagicTryOn at comparable inference cost.The comparison includes both quality metrics and WildVVT inference time and GPU memory usage.Source: §4.3 · p. 7, Table 3 · p. 8

Ablation Study

• Removing GSKS or Frame-Shared 3D-RoPE degrades performance, particularly garment consistency, while separate positional encodings weaken garment-detail transfer.The ablation reports incorrect shoulder straps and missing white chest stripes without Frame-Shared 3D-RoPE.Source: §4.4 · p. 8
• With only 500 video-level pseudo samples, the full model reaches performance comparable to the variant without Stage 2 trained with 2,500 samples.Stage 2 improves mask-free localization and fine-tuning efficiency across all tested pseudo-video scales.Source: §4.4 · p. 8, Figure 7 · p. 8

Conclusion

• BooM-VVT learns mask-free try-on localization from image-level pseudo data, reducing reliance on costly video-level pseudo data.Its multi-stage training strategy supports mask-free video virtual try-on with minimal video-level pseudo data.Source: §5 · p. 9
• GSKS selects garment-relevant keyframes, while Frame-Shared 3D-RoPE establishes spatiotemporal correspondences for garment appearance transfer.These components target garment consistency and fine-grained detail preservation.Source: §5 · p. 9
• OmniView provides multi-view garment references and broader category coverage for challenging viewpoints and diverse try-on tasks.The dataset supports layered, cross-category, and multi-view try-on settings.Source: §5 · p. 9
• Extensive experiments demonstrate superior temporal consistency and garment fidelity under unconstrained real-world scenarios.The reported conclusion covers the framework's overall experimental performance.Source: §5 · p. 9

• Image try-on methods use warping-and-fusion or diffusion-based reference injection, but dense preprocessing remains common.Source: §A.1 · p. 11
• Video try-on must preserve temporal coherence, motivating temporal modules, garment encoders, memory refinement, and video DiT architectures.Source: §A.2 · p. 11
• Mask-based try-on offers spatial controllability but depends on mask quality, whereas mask-free methods implicitly localize try-on regions from data.Source: §B.1 · p. 11, §B.1.1 · p. 11
• Keyframe-driven methods decompose VVT into keyframe try-on and keyframe-guided video generation, enabling use of unpaired videos and pretrained backbones.Source: §B.1.2 · p. 11
• Existing keyframe-driven methods still rely on explicit masks, limiting robustness in challenging real-world scenarios.Source: §B.1.2 · p. 12
• GSKS selects garment-relevant limbs and body regions using DWPose and SAM, with K=2, λ=0.2, and Δmin = 4 frames in the default setting.Source: §C.1 · p. 12

• The VLM protocol evaluates Appearance & Background Consistency, Garment Consistency, and Overall Visual Quality with Gemini 3.1 Pro.Source: §C.2 · p. 12
• Each VLM dimension is scored independently from 0 to 5 using the source video, target garment image, and generated try-on video.Source: §C.2.1 · p. 12
• A human study found strong agreement between Gemini-based and human judgments, supporting the reliability of the VLM evaluation.Source: §C.2.1 · p. 12
• VFID is included for comparability with prior work but is not used as the primary metric in the main paper.Source: §C.2.1 · p. 12
• The open-source VFID toolkit evaluates only the first 10 frames, so it cannot fully assess videos with later motion or content changes.Source: §C.2.2 · p. 12

• The strategy uses image-level pseudo data for Stage 2 and video-level pseudo data for Stage 3.Source: §C.3 · p. 13
• Qwen3-VL-32B analyzes each person image and generates a garment-editing instruction for pseudo-image construction.Source: §C.3 · pp. 13–14
• GSKS determines garment-related limb sets and body regions before scoring and selecting informative, complementary keyframes.Source: §C.3 · p. 14
• Qwen-Image-Edit-2511 modifies the garment region, while Qwen3-VL-32B verifies instruction consistency and rejects failed samples.Source: §C.3 · p. 14
• The pipeline constructs 30K pseudo images spanning diverse garment types, body poses, and real-world backgrounds.Source: §C.3 · p. 14
• Video-level pseudo try-on videos are synthesized with the Stage 1 BooM-VVT model because existing open-source video editors lack sufficient quality.Source: §C.3 · p. 14

• BooM-VVT achieves the best corrected-VFID results in three of four ViViD and WildVVT settings and ranks second in the remaining setting.Source: §D.1 · p. 14
• On WildVVT, BooM-VVT achieves the lowest VFIDI by a clear margin, indicating stronger generalization in unconstrained real-world scenarios.Source: §D.1 · p. 14
• Human evaluation on WildVVT uses 10 evaluators, the same ABC, GC, and OVQ dimensions, and a 0–5 rating scale.Source: §D.1 · p. 14

• BooM-VVT achieves the highest human evaluation scores across ABC, GC, and OVQ.ABC measures Appearance & Background Consistency, GC measures Garment Consistency, and OVQ measures Overall Visual Quality on a 0–5 scale.Source: §D.1.1 · p. 14, Table 7 · p. 15
• Pearson correlation of approximately r=0.9 indicates strong agreement between human scores and Gemini-based evaluation.Source: §D.1.1 · p. 14
• Human evaluation uses Appearance & Background Consistency, Garment Consistency, and Overall Visual Quality as its three dimensions.Evaluators score each dimension from 0 to 5, with higher scores indicating better quality.Source: Table 7 · p. 15

Large motion

• Under large motion, BooM-VVT produces more stable garment appearance and better temporal consistency than existing methods.Its full mask-free model also outperforms the mask-dependent BooM-VVT (Stage 1) in garment consistency, indicating that inaccurate fast-motion masks can degrade localization.Source: §D.2 · p. 15

Severe occlusion

• Under severe occlusion, BooM-VVT maintains plausible garment appearance despite large portions of the body being occluded.Existing methods often produce obvious artifacts or incorrect garment structures in this setting.Source: §D.2 · p. 15

Object interaction

• During object interaction, BooM-VVT better preserves person–object relationships than existing methods.Other methods may introduce artifacts in interaction regions or distort and erase interacting objects.Source: §D.2 · p. 15

Diverse try-on tasks

• OmniView enables layered, cross-category, and multi-view try-on, demonstrating generalization across diverse try-on scenarios.The settings use sequential layer-by-layer dressing, text-guided cross-category information, and garment images from different viewpoints.Source: §D.3 · p. 15

Garment-Sensitive Keyframe Sampling

• GSKS selects keyframes using garment-relevant regions and complementary viewpoint coverage, improving fine-grained garment-detail preservation.Compared with global pose-difference sampling, it supports more faithful logos, patterns, and other garment details in keyframe images and final videos.Source: §D.4.1 · pp. 15–16

OmniView ablation

• OmniView improves cross-view consistency in keyframe try-on results.Without OmniView, garment details such as belts can appear in one viewpoint and disappear in another.Source: §D.4.2 · p. 16, §D.4.1 · p. 16
• OmniView improves handling of multi-view garment inputs by preserving sharper textures and more accurate garment structures.The model trained without OmniView can generate incorrect back-view textures even when front- and back-view references are provided.Source: §D.4.2 · p. 16

Weighting-factor sensitivity

• When λ is between 0.2 and 0.6, SSIM stays around 0.871 and LPIPS around 0.062 on WildVVT.Performance remains stable across a wide range of λ values, while overly large values cause slight degradation.Source: §D.4.3 · p. 16, §D.4.2 · p. 16, Table 8 · p. 16

Evaluation protocol

• The VLM-based evaluation scores Appearance & Background Consistency, Garment Consistency, and Overall Visual Quality on integer scales from 0 to 5.Source: Figure 11 · p. 17

Qualitative evaluations

• Qualitative evaluations cover large motion, severe occlusion, object interaction, and multiple try-on task configurations.The figures include layered, cross-category, and multi-view try-on, plus single-view and front-and-back multi-view garment inputs.Source: Figure 12 · p. 18, Figure 13 · p. 18, Figure 14 · p. 19