TexTailor: Texture-Preserving Video Virtual Try-On via Adaptive Garment Conditioning

Zijing Qin; Jun Zhou; Ruicheng Zhang; Jiaqi Hou; Zunnan Xu; Ronghui Li; Zhenyu Xie; Xiu Li

cs.CVSep 30, 2026

TexTailor: Texture-Preserving Video Virtual Try-On via Adaptive Garment Conditioning

Authors: Zijing Qin, Jun Zhou, Ruicheng Zhang, Jiaqi Hou, Zunnan Xu, Ronghui Li, Zhenyu Xie, Xiu Li

Organizations: Tsinghua University, China · Mohamed bin Zayed University of Artificial Intelligence, UAE

Abstract

Video virtual try-on has attracted increasing attention due to its broad potential in digital fashion and intelligent e-commerce. However, existing methods primarily focus on low-resolution settings and still face substantial challenges when extended to high-resolution scenarios. These limitations can be attributed to two main factors: (1) the insufficient utilization of rich garment reference information, and (2) the lack of explicit positional modeling between garment and video representations during cross-modal interaction, which weakens fine-grained local correspondence. To address these issues, we propose TexTailor, a high-fidelity video virtual try-on framework built upon a pretrained video Diffusion Transformer. Specifically, we introduce a timestep-adaptive modulation mechanism to dynamically adjust garment visual representations throughout denoising. We further develop a frame-aligned positional encoding strategy to strengthen garment-to-video correspondence, together with a multi-source injection design that reduces interference among heterogeneous conditions. Extensive experiments on multiple video virtual try-on benchmarks, including the high-resolution Eevee dataset, demonstrate that TexTailor achieves competitive performance in garment detail preservation, temporal consistency, and overall video quality.

Figures & tables

Figure 1: TexTailor generates high-fidelity video virtual try-on results with temporally consistent motion while preserving garment textures, patterns, and local structures across diverse poses and viewpoints.

Figure 2: Overview of TexTailor. (a) A DiT backbone integrates structural input, text token, visual token, and garment latent token through MCAI for video try-on generation. (b) TAVM modulates visual features with timestep-adaptive token-wise gating. (c) FAC-RoPE assigns frame-aligned temporal coordinates to static garment latent tokens, preserving native 3D RoPE while establishing stable garment-to-video spatial correspondence.

Figure 3: Qualitative comparison on the high-resolution Eevee benchmark. Compared with existing methods, TexTailor better preserves garment textures, patterns, and local structures while maintaining temporally coherent appearance across frames.

Method	Full-shot			Close-up			GPU Mem.	Time
Method	VFID $\_R$ $\downarrow$	VFID $\_I$ $\downarrow$	VGID $\uparrow$	VFID $\_R$ $\downarrow$	VFID $\_I$ $\downarrow$	VGID $\uparrow$	GPU Mem.	Time
ViViD ( Fang et al. 2024 )	0.565	12.859	0.514	1.253	12.665	0.533	72.21G	348.96s
MagicTryOn ( Li et al. 2025b )	0.187	9.783	0.512	0.752	11.285	0.538	69.54G	589.73s
CatV 2 TON ( Chong et al. 2025 )	0.751	9.086	0.520	0.673	11.935	0.522	40.18G	357.21s
TexTailor	0.149	8.872	0.531	0.489	10.917	0.543	30.32G	230.72s

Table 1: Quantitative comparison on the Eevee benchmark at (

1088\times 816

). We report results under both full-shot and close-up settings, together with GPU memory usage and inference time. The best and second-best results are highlighted in bold and underline , respectively.

Method	VFID ${}_{I}^{p}$ $\downarrow$	VFID ${}_{R}^{p}$ $\downarrow$	SSIM $\uparrow$	LPIPS $\downarrow$	VFID ${}_{I}^{u}$ $\downarrow$	VFID ${}_{R}^{u}$ $\downarrow$	GPU Mem.	Time
ViViD ( Fang et al. 2024 )	17.1847	0.6382	0.8041	0.1216	21.6925	0.8347	62.59G	204.183s
CatV 2 TON ( Chong et al. 2025 )	13.4821	0.2876	0.8742	0.0647	19.3846	0.5179	27.66G	209.127s
MagicTryOn ( Li et al. 2025b )	8.3165	0.2298	0.9024	0.0598	14.6032	0.3137	51.51G	345.271s
TexTailor	7.9428	0.2645	0.9017	0.0619	13.8754	0.2861	25.32G	196.852s

Table 2: Quantitative comparison on the ViViD benchmark. We report paired and unpaired evaluation metrics, together with GPU memory usage and inference time. The best and second-best results are highlighted in bold and underline , respectively.

Setting	Metric	w/o TAVM-S	w/o TAVM-D	w/o TAVM-G	w/o TAVM	w/o FAC	w/o MCAI-T	w/o MCAI-G	w/o Mask	Full Model
Full-shot	VFID R $\downarrow$	0.184	0.181	0.176	0.224	0.207	0.179	0.201	0.198	0.149
	VFID I $\downarrow$	9.512	9.463	9.382	10.024	9.781	9.427	9.694	9.821	8.872
	VGID $\uparrow$	0.520	0.521	0.522	0.512	0.515	0.523	0.516	0.517	0.531
Close-up	VFID R $\downarrow$	0.551	0.543	0.535	0.612	0.581	0.529	0.568	0.592	0.489
	VFID I $\downarrow$	11.612	11.534	11.462	12.146	11.873	11.421	11.754	11.982	10.917
	VGID $\uparrow$	0.542	0.535	0.526	0.514	0.539	0.527	0.540	0.538	0.543

Table 3: Ablation study on the Eevee benchmark. We evaluate the contribution of each component in TexTailor.

Figure 4: Visualization of timestep-wise token gating in TAVM. (a) Gate activation maps across different denoising steps. (b) Representative token regions selected from the garment reference. (c) Gate evolution curves of the selected token regions in (b). TAVM progressively shifts its emphasis from coarse garment structure to fine-grained local details during denoising.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Full-shot			Close-up			GPU Mem.	Time
Method	VFID R $\downarrow$	VFID I $\downarrow$	VGID $\uparrow$	VFID R $\downarrow$	VFID I $\downarrow$	VGID $\uparrow$	GPU Mem.	Time
ViViD ( Fang et al. 2024 )	0.389	12.194	0.506	0.936	12.198	0.533	64.12G	216.734s
MagicTryOn ( Li et al. 2025b )	0.161	9.865	0.520	0.595	11.262	0.534	49.83G	333.906s
CatV 2 TON ( Chong et al. 2025 )	0.746	9.141	0.518	0.632	11.847	0.538	29.14G	197.683s
TexTailor	0.153	8.996	0.529	0.497	11.071	0.553	24.87G	188.436s

Appendix

Table 4: Quantitative comparison on the Eevee benchmark at

832\times 624

. We report results under both full-shot and close-up settings, together with GPU memory usage and inference time. The best and second-best results are highlighted in bold and underline , respectively.

Figure 5: Qualitative comparison on challenging garment cases. TexTailor better preserves complex patterns, fine-grained textures, and local garment structures compared with existing methods.

Figure 6: Qualitative comparison across diverse garment categories. TexTailor achieves more faithful garment transfer with improved texture preservation and structural consistency.

Figure 7: Additional results of TexTailor on diverse garment categories, including sweaters, dresses, and pants.

Figure 8: Additional results of TexTailor on garments with complex patterns and structures.

Figure 9: Additional results demonstrating the generalization ability of TexTailor across various garments and motion scenarios.

Figure 10: Qualitative ablation comparison of TexTailor. Removing individual components leads to degraded garment fidelity, including texture distortion, blurred details, and weaker pattern preservation. The full model achieves the most faithful garment appearance.

Explore similar work

UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on
Aug 6, 2026Yushe Cao, Shikun Feng, Fei Shen +5Virtual Try-On Garments
Tsinghua University · Zhongguancun Academy · National University of Singapore +2
iTryOn: Mastering Interactive Video Virtual Try-On with Spatial-Semantic Guidance
May 20, 2026Jun Zheng, Zhengze Xu, Mengting Chen +6Virtual Try-On Garments
Shenzhen Campus of Sun Yat-sen University · Taobao & Tmall Group of Alibaba
OmniTryOn: Video Try-On Anything at Once!
Jun 7, 2026Changliang Xia, Chengyou Jia, Minnan Luo +3Virtual Try-On Garments
Xi’an Jiaotong University Xi’an, China

Aug 6, 2026, cs.CVJ/K move · Enter open · S save

UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on

Yushe Cao, Shikun Feng, Fei Shen, Haikuo Peng+4

Tsinghua University · Zhongguancun Academy · National University of Singapore+2

Virtual Try-On Garments Video Generation Geometric Priors

Open paper AskPDF arXiv

Video Virtual Try-On (VVT) synthesizes a video of a person wearing a target garment while preserving identity, motion, and scene dynamics. Dominant approaches cast VVT as mask-conditioned video inpainting and rely on separate modules for human parsing, pose estimation, and garment warping. This multi-stage design complicates deployment and, more critically, allows errors in explicit geometric priors to propagate irreversibly into the generated video. We present UniVVT, a unified end-to-end framework that reframes VVT as semantically conditioned video generation, eliminating mask, pose, and warping modules at inference. At its core, a scene-task perceiver built on a Multimodal Large Language Model jointly encodes the source video, target garment, and task instruction into compact, task-aware latent tokens, implicitly capturing what to transfer and where and how to transfer it. A lightweight semantic bridge then aligns these tokens with the conditioning space of a diffusion-based video generator, enabling coherent garment transfer. To robustly couple the heterogeneous components, we devise a three-stage progressive training strategy comprising semantic alignment, joint task adaptation, and flexible-resolution refinement. Extensive experiments demonstrate that UniVVT achieves state-of-the-art performance across multiple benchmarks, validating implicit semantic guidance as a simple and effective alternative to fragile geometric preprocessing for end-to-end virtual try-on.

Abstract

Explore similar work

UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on

iTryOn: Mastering Interactive Video Virtual Try-On with Spatial-Semantic Guidance

OmniTryOn: Video Try-On Anything at Once!