PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.39335v1cs.CV

TL;DR

Existing video try-on methods largely target low-resolution settings and leave garment conditioning and garment-to-video positional alignment gaps. TexTailor combines adaptive visual modulation, frame-aligned positioning, and multi-source injection, achieving competitive results across multiple benchmarks in garment detail, temporal consistency, and video quality.

  • Problem

    Existing video try-on methods largely target low-resolution data, while fine-grained garment conditioning and explicit garment-to-video positional correspondence remain limited.

    Source: Introduction · p. 1

  • Method

    TexTailor uses a pretrained video Diffusion Transformer with timestep-adaptive visual modulation, frame-aligned garment-to-video positioning, and decoupled multi-source conditioning.

    Source: Introduction · p. 1

  • Results

    TexTailor achieves competitive performance in garment-detail preservation, temporal consistency, and overall video quality across multiple benchmarks, including high-resolution Eevee.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The results highlight the importance of adaptive and spatially aligned garment conditioning for high-fidelity video virtual try-on.

    Source: Conclusion · p. 7

    The framework relies on auxiliary pose, agnostic-representation, and video-mask inputs, and large deformation, heavy occlusion, and major viewpoint changes remain difficult scenarios.

    Source: Discussion and Limitations · p. 11

Abstract

from arXiv · show

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.

Introduction

High-resolution video try-on requires preserving fine garment detail while maintaining coherent motion, but existing conditioning and positional modeling leave gaps. TexTailor combines adaptive garment conditioning and frame-aligned alignment, with competitive results across multiple benchmarks.

Introduction

  • Video try-on must preserve garment shapes, textures, and patterns while maintaining temporal consistency of human identity and motion.Unlike image-based try-on, it must satisfy these appearance and motion requirements across continuous movement and viewpoint changes.Source: Introduction · p. 1
  • Existing methods are mostly developed at low resolution and often use static garment features, lack explicit garment-to-video positional correspondence, and fuse heterogeneous conditions in shared attention spaces.These limitations are especially relevant at high resolution, where textures, stitching, patterns, and material structures are vulnerable to blurring, loss, and temporal drift.Source: Introduction · p. 1
  • TexTailor builds on a pretrained video Diffusion Transformer with timestep-adaptive visual modulation, frame-aligned garment-to-video positioning, and decoupled multi-source conditioning.MCAI processes textual cues, timestep-modulated visual features, and frame-aligned garment latents through parallel cross-attention pathways.Source: Introduction · pp. 1–2
  • TexTailor achieves competitive performance in garment-detail preservation, temporal consistency, and overall video quality across multiple benchmarks, including high-resolution Eevee.The paper describes consistent improvements in garment fidelity, temporal consistency, and high-resolution generation quality compared with existing approaches.Source: Introduction · p. 2

Related Work

Video generation has advanced in visual realism and temporal coherence, but virtual try-on adds the challenge of preserving garment-specific appearance through motion. Prior video try-on work has evolved toward diffusion and DiT models, while high-resolution fine-detail conditioning remains underexplored.

Video Generation

  • Video generation methods improve realism through temporal attention, reference guidance, and motion-specific modules, but try-on also requires garment fidelity and plausible clothing interaction with human motion.The passage argues that generic video-generation models are often insufficient without task-specific garment modeling and conditioning.Source: Video Generation · p. 2

Video Virtual Try-On

Method

TexTailor conditions a pretrained video Diffusion Transformer with adaptive garment features, frame-aligned positional encoding, and disentangled cross-attention, and adds garment-mask weighting during training.

Preliminary

  • The pretrained video DiT uses VAE video latents and timestep-weighted flow matching, conditioned on person and motion structure alongside garment-reference features.Structural conditions include person video, pose video, agnostic video, and clothing mask.Source: Preliminary · pp. 3–4

Timestep-Adaptive Visual Modulation

  • TAVM adapts garment visual features and token-wise gates to each denoising timestep, allocating garment information for coarse-to-fine reconstruction.SigLIP and DINO visual tokens form the garment condition injected through multi-source cross-attention.Source: Timestep-Adaptive Visual Modulation · p. 4

Frame-Aligned 3D Garment Cross-RoPE

  • FAC-RoPE assigns garment keys the video query’s frame coordinate, establishing positional correspondence without artificial temporal bias.The design is intended to improve local alignment and cross-frame texture consistency.Source: Frame-Aligned 3D Garment Cross-RoPE · p. 4

Multi-Source Cross-Attention Injection

  • MCAI injects text semantics, modulated visual features, and garment latent tokens through separate cross-attention branches to reduce interference among conditions.Independent projection and attention let the modalities contribute complementary information while retaining their semantic characteristics.Source: Multi-Source Cross-Attention Injection · p. 4

Training Objective

  • A mask-aware loss reweights the flow-matching objective with the mask video to strengthen training signals in garment regions while retaining the original objective.The strength of mask-aware supervision is controlled by λ.Source: Training Objective · p. 4

Experiments

TexTailor is competitive across high-resolution Eevee and conventional ViViD evaluations, with qualitative comparisons and component ablations supporting its garment-detail and temporal-consistency goals.

Experimental Setup

  • Eevee provides 1088×816 videos and 2400×1800 garment images; evaluation covers paired and unpaired settings on Eevee and ViViD.Training uses 49-frame samples and two resolution stages, ending at 1088 × 816; inference uses 25 denoising steps.Source: Experimental Setup · pp. 4–5

Quantitative Comparison

  • TexTailor achieves competitive results on high-resolution Eevee and ViViD, with lower GPU memory consumption and faster inference than existing methods.The comparison covers Eevee full-shot and close-up settings and ViViD; Table 1 and Table 2 report quantitative comparisons.Source: Table 1 · p. 5, Quantitative Comparison · p. 6, Table 2 · p. 6

Qualitative Comparison

  • On complex garments and varied poses, TexTailor better retains collars, textures, patterns, and local structures while keeping appearance more coherent across frames.Figure 3 presents qualitative comparisons on high-resolution Eevee.Source: Figure 3 · p. 5, Qualitative Comparison · p. 6

Ablation Study

  • Removing either TAVM visual stream or its timestep-aware gating degrades performance; its gates shift emphasis from garment structure early to fine detail later.The token-region curves show structure-related regions more active early and detail-related regions gaining importance during later refinement.Source: Ablation Study · pp. 6–7, Figure 4 · p. 7
  • Removing FAC-RoPE degrades performance; without frame-aligned positional encoding, garment tokens lack consistent spatial anchors with video tokens.The passage links this alignment to accurate garment-region correspondence during generation.Source: Ablation Study · p. 7
  • Removing either MCAI branch degrades performance, indicating complementary information from semantic guidance and spatially aligned garment features.MCAI injects the condition sources through disentangled cross-attention branches.Source: Ablation Study · p. 7
  • Removing mask-aware loss degrades performance; garment-region-focused supervision provides stronger optimization signals for reconstructing local garment details.The cited study identifies textures, complex patterns, and subtle structures as the targeted details.Source: Ablation Study · p. 7

Conclusion

TexTailor combines adaptive garment guidance, frame-aligned positional correspondence, and multi-source conditioning for high-fidelity video try-on. Experiments across benchmarks, including high-resolution Eevee, show better preservation of local garment details.

Conclusion

  • TexTailor combines timestep-adaptive visual modulation, frame-aligned 3D garment Cross-RoPE, and multi-source cross-attention injection to preserve fine garment details in high-resolution video try-on.The paper presents these designs as adaptive and spatially aligned garment conditioning for high-fidelity generation.Source: Conclusion · p. 7

Supplementary Materials

The supplement extends the evaluation with additional quantitative and qualitative results, ablations, and a discussion of framework limitations.

Supplementary Materials

  • The supplement adds resolution-specific quantitative results, qualitative comparisons and visualizations, further ablations, and a limitations discussion.Source: Supplementary Materials · p. 11

More Quantitative Results

  • At 832×624, TexTailor maintains competitive performance in full-shot and close-up settings while using less GPU memory and inferring faster than existing approaches.Table 4 reports both settings along with GPU memory usage and inference time.Source: More Quantitative Results · p. 11, Table 4 · p. 12

More Qualitative Results

More Ablation Results

  • Removing individual components degrades garment fidelity through texture distortion, blurred local details, and weaker pattern preservation; the full model is more faithful.The qualitative ablations are reported as consistent with the quantitative ablation results.Source: More Ablation Results · p. 11, Figure 10 · p. 14

Discussion and Limitations

  • The framework still relies on pose information, agnostic representations, and video masks, while large deformation, heavy occlusion, and viewpoint changes remain difficult.The paper identifies simplified conditioning and improved robustness under complex motion and viewpoint variations as future directions.Source: Discussion and Limitations · p. 11
Loading 2609.39335v1…