ECCV 2026 · Video Virtual Try-On

TripVVT

A large-scale triplet dataset and a coarse-mask baseline for in-the-wild video virtual try-on.

Paper Dataset GitHub

10,031: high-resolution triplets
100: benchmark cases
Coarse: human-mask prior

In-the-wild result video

Garment + sourceTripVVT output

Dingbao Shao^1,*Song Wu^2,*Shenyi Wang¹Ye Wang³Ziheng Tang¹Fei Liu⁴Jiang Lin¹Xinyu Chen¹Qian Wang²Ying Tai¹Jian Yang¹Zili Yi^1,†

¹Nanjing University²JIUTIAN Research³Jilin University⁴ByteDance Inc.

^*Equal contribution^†Corresponding author

Results in motion

Try-on beyond the studio.

Real-world motion, cluttered backgrounds, and diverse garments—all evaluated as videos, not isolated frames.

Real-world sceneCross-garment transfer

In-the-wild motionTemporal consistency

VIVID-S-TestGarment fidelity

Overview

Training data that matches the real task.

TripVVT connects a large-scale triplet dataset, a stable spatial prior, and a challenging video benchmark in one research stack.

Video virtual try-on still struggles outside curated scenes because large-scale in-the-wild triplet supervision is scarce and fine-grained garment masks are fragile under motion, occlusion, and clutter.

We introduce TripVVT-10K, a dataset with explicit video-level cross-garment supervision; TripVVT, a Diffusion Transformer baseline that replaces garment masks with a stable coarse human-mask prior; and TripVVT-Bench, a 100-case evaluation suite spanning video quality, try-on fidelity, background consistency, and temporal coherence.

Across academic and commercial comparisons, TripVVT improves video quality and garment fidelity while generalizing more reliably to unconstrained videos. The dataset and benchmark are publicly released to support reproducible progress.

01 · Dataset

From wild videos to supervised triplets.

TripVVT-10K turns unconstrained source videos into aligned training triplets while preserving motion and scene diversity.

TripVVT-10K data construction pipeline — **TripVVT-10K construction.** A garment-swapped anchor frame and pose sequence produce a synthesized original video; image-to-garment synthesis creates the canonical garment reference; the raw source video provides the try-on target.

02 · Method

A coarse prior. A stronger baseline.

A human-mask prior preserves the scene without depending on brittle garment-level segmentation in every frame.

TripVVT model architecture — **TripVVT architecture.** The DiT backbone consumes the original video, pose, and coarse human mask, while garment and line encoders inject appearance and structural cues into the transformer blocks.

Inside TripVVT-10K

Triplets, not isolated frames.

Each sample pairs a canonical garment with an original video and its cross-garment try-on target.

03 · Results

Qualitative evaluation across benchmarks.

Compare garment fidelity, background preservation, and temporal consistency on TripVVT-Bench and VIVID-S-Test.

Limitations

Where the model still breaks.

Challenging motion, occlusion, and garment structures remain open problems. We include representative failures rather than hiding them.

Citation

Build on TripVVT.

If the dataset, benchmark, or baseline supports your work, please cite the paper.

@article{shao2026tripvvt,
  title   = {TripVVT: A Large-Scale Triplet Dataset and a Coarse-Mask Baseline for In-the-Wild Video Virtual Try-On},
  author  = {Shao, Dingbao and Wu, Song and Wang, Shenyi and Wang, Ye and Tang, Ziheng and Liu, Fei and Lin, Jiang and Chen, Xinyu and Wang, Qian and Tai, Ying and Yang, Jian and Yi, Zili},
  journal = {arXiv preprint arXiv:2604.27958},
  year    = {2026}
}