TripVVT-10K: In-the-Wild Video Virtual Try-On

• The paper introduces TripVVT-10K as the first large-scale, high-resolution in-the-wild video triplet dataset that provides explicit video-level cross-garment supervision.
• It integrates synchronized auxiliary modalities such as human-mask videos, pose videos, garment line maps, and text prompts to guide realistic garment transfer.
• Benchmark evaluations reveal that TripVVT-10K significantly enhances video quality, garment fidelity, and temporal coherence over previous paired virtual try-on datasets.

TripVVT-10K is a video virtual try-on dataset introduced in "TripVVT: A Large-Scale Triplet Dataset and a Coarse-Mask Baseline for In-the-Wild Video Virtual Try-On" (Shao et al., 30 Apr 2026). It is presented as the first large-scale, high-resolution in-the-wild video triplet dataset for virtual try-on and, in the paper’s abstract, as the largest and most diverse in-the-wild triplet dataset to date. Its central purpose is to provide the supervision regime that prior video virtual try-on resources lacked: explicit video-level cross-garment supervision in unconstrained real-world scenes. Each sample is organized as a triplet of (original video, garment reference, try-on video), supplemented by synchronized auxiliary modalities used by the associated TripVVT framework. The dataset is both a training resource and the basis for TripVVT-Bench, a held-out benchmark for evaluating video quality, garment fidelity, background preservation, and temporal coherence in in-the-wild settings (Shao et al., 30 Apr 2026).

1. Position within video virtual try-on research

TripVVT-10K addresses a specific bottleneck in video virtual try-on: prior public video datasets were described as studio-like, low-diversity, low-resolution, or limited to pairwise supervision rather than true cross-garment supervision over time. The paper argues that this mismatch impedes realistic deployment, especially in outdoor or cluttered scenes with pose changes, occlusions, lighting variation, and complex backgrounds. In that context, the dataset is designed as a supervision source for unconstrained, high-resolution, temporally coherent garment transfer rather than for narrowly controlled laboratory conditions (Shao et al., 30 Apr 2026).

The term large-scale in-the-wild triplet dataset is used in a technically specific sense. “Large-scale” refers to 10,031 triplets built from approximately 20K web-crawled in-the-wild human videos. “In-the-wild” refers to unconstrained web videos covering outdoor and real-world environments, low light, seated poses, crowded public environments such as subway interiors, multiple people, varied viewpoints, and unconstrained motion. “Triplet dataset” means that each example contains (original video, garment reference, try-on video), enabling supervision across garments for the same person and motion sequence.

A central distinction is the dataset’s video-level cross-garment supervision. Existing video datasets such as VVT and ViViD are described as paired, but not triplet-based in a way that links a person/video in one garment to the same person/video in another garment plus a garment reference. TripVVT-10K instead provides an input video of a person wearing one outfit, a garment reference image, and a target try-on video of the same person sequence wearing another outfit. This supports learning to transfer garments across videos while preserving identity, pose, motion, and background.

The comparison reported in Table 1 situates TripVVT-10K among image and video virtual try-on resources.

Dataset	Format and supervision	Scene/person/resolution/samples
VITON-HD	image, paired	studio, single-person, 768×1024, 11,647
DressCode	image, paired	studio, single-person, 768×1024, 53,792
Street TryOn	image, pseudo pairing	outdoor, single-person, variable resolution, 14,453
LAION-G.	image, triplet	outdoor, single-person, variable resolution, 60,000
VVT	video, paired	studio, single-person, 192×256, sample count not printed
ViViD	video, paired	studio, single-person, 624×832, 9,700
TripVVT-10K	video, triplet	studio and outdoor, single and multiple persons, 720×1280, 10,031

This comparison clarifies an important boundary condition. TripVVT-10K is not the largest dataset across all virtual try-on modalities, because LAION-G. has 60,000 image triplets. Its claim is narrower: it is positioned as the largest and most diverse in-the-wild video triplet dataset.

2. Dataset structure, scale, and supervision semantics

TripVVT-10K contains 10,031 triplets, covers 30 clothing categories, normalizes all videos to 720 × 1280 resolution, and is built from approximately 20K crawled in-the-wild human videos. During preprocessing, each source video is uniformly subsampled to 121 frames. During model optimization, training uses 49-frame clips sampled from the dataset. The associated benchmark contains 100 fully held-out triplets. Table 1 further states that the dataset includes both outdoor and studio scenes, contains single-person and multiple-person videos, and uses Triplet pairing (Shao et al., 30 Apr 2026).

The triplet definition is unusual because of the paper’s reverse design. A sample consists of (original video, garment reference, try-on video), but the original video is not the raw source clip. Instead, it is the synthesized garment-swapped sequence used as model input. The try-on video is the authentic source video and serves as ground truth. The paper states this explicitly: “Following a reverse design, we treat the synthesized garment-swapped video as the model input (original video), while the authentic source video is used as the ground-truth try-on video.” This naming convention is central to interpreting the dataset correctly.

Garment coverage spans 30 types of clothing categories from upper body, lower body, and full body. The representative category names shown in the paper include upper-body garments such as t-shirt, cami tank, shirt/blouse, sweater knit, jacket, sweatshirt, blazer/suit jacket, hoodie, cardigan; lower-body garments such as jeans, trousers, leggings, shorts, wide leg pants, skirt, midi skirt; and full-body garments such as dress, long dress, romper, qipao or ethnic dress, other full body. The text emphasizes breadth of category coverage, but exact per-category counts are not provided.

Each sample also includes several auxiliary modalities used during training:

1. a human-mask video
2. a pose video estimated by DWPose from the original video
3. a garment line map, extracted from the garment reference using AniLines
4. a text prompt describing the overall scene and the target person’s outfit over the entire video, generated by Qwen-VL-Max

The paper is equally explicit about what is not provided. It does not claim to include human parsing maps, DensePose, tracking labels, explicit identity annotations, manual garment metadata, person IDs, garment IDs, category labels as downloadable metadata, or bounding-box tracks. It also does not provide exact counts for identities, garments, category-wise distribution, total duration, total frame count, or geographic provenance.

3. Construction pipeline and reverse-training synthesis

TripVVT-10K is constructed through an automated synthesis pipeline rather than by directly collecting naturally paired real-world videos of the same person wearing different clothes (Shao et al., 30 Apr 2026). The process begins with approximately 20K web-crawled in-the-wild human videos. These are normalized to a fixed resolution, low-quality or heavily compressed clips are removed, and each surviving video is uniformly subsampled to 121 frames.

For each source clip, the authors use Gemini-2.5-Flash together with SAM2 to generate a temporally consistent human-mask video. This human-mask video serves a dual role. It supports downstream dataset construction, and it later becomes the principal spatial prior used by the TripVVT model.

The synthetic triplet is then created through the paper’s reverse-training paradigm. Starting from a real source video, the authors create a garment-swapped synthetic sequence that becomes the dataset’s original video, while the real source video is retained as the try-on video. This arrangement allows training against a real target sequence rather than a synthetic target.

The garment-swapping process is multi-stage. The garment region on the first frame is segmented using SegFormer on the cropped human and then propagated through time with SEC. For first-frame editing, a garment is randomly sampled from DressCode, and Gemini-2.5-Flash produces a garment-specific textual description. The subject crop from the first source frame and this description are then fed into Nano Banana to synthesize a high-quality swapped first frame. Using ViTPose to extract the pose sequence from the source video, the authors create a masked source video and call Wan-Animate to generate a temporally coherent garment-swapped sequence conditioned on the swapped first frame, the pose sequence, and the masked source video. That output becomes the dataset’s original video.

The garment reference image is reconstructed from the clothing worn in the try-on video. The person is extracted from the first source frame using the human mask, clutter and other people are removed with RMBG-2.0, and Nano Banana is used again to produce a canonical product-style, background-free, front-view garment image.

Quality control operates in two stages. First, Gemini-2.5-Flash automatically evaluates garment-reference fidelity and consistency between the synthetic original video and the source video, checking attributes such as color, texture, and structure. Triplets with poor garment fidelity or severe artifacts are discarded. Second, the remaining candidates are manually reviewed to remove identity shifts, temporal flicker, and visible editing artifacts. The final curated result is TripVVT-10K.

4. Coupling to the TripVVT framework

The dataset is not an isolated resource; it is the supervision substrate for the TripVVT method. TripVVT is described as a Diffusion Transformer-based end-to-end video virtual try-on framework built on wan2.1-Fun-14B-control, itself fine-tuned from Wan2.1-I2V-14B, and reusing the DiT-based backbone of MagicTryOn (Shao et al., 30 Apr 2026). Its conditioning design is aligned closely with the modalities provided by TripVVT-10K.

The model takes the synchronized original video, human-mask video, and pose video of the target person as inputs. It also consumes the reference garment image, garment line map, and a global text prompt describing the scene. A VAE encoder extracts spatio-temporal features from the original video, which are concatenated with resized human-mask features. The reference garment image passes through a garment encoder to produce garment tokens, the line map is processed by a line encoder, and a text encoder supplies conditioning from the global prompt. These features are fed into a DiT backbone with what the figure labels Magic-tryon attention blocks, and a VAE decoder produces the output video.

The paper’s principal methodological claim is the value of the human-mask prior. Garment-mask conditioning is characterized as precise but fragile, while mask-free approaches are characterized as flexible but liable to alter unrelated content. The proposed compromise is to condition on a simple, temporally consistent human mask. This supplies a robust spatial prior that constrains editing to the person region while avoiding dependence on fine-grained garment segmentation. TripVVT-10K is therefore not only a collection of triplets; it is also a dataset whose auxiliary modalities are tailored to a coarse-mask formulation.

Training details make that dependence concrete. Training follows a three-stage schedule:

• Stage 1: use the TripVVT supplementary set only, for 25k steps at 512×384
• Stage 2: add TripVVT-10K, then train another 25k steps at mixed resolutions 512×384 and 592×334
• Stage 3: increase resolution to 832×624 and 960×544 for 5.5k steps

All stages use 49-frame clips, batch size 2, AdamW, and a constant learning rate of 1 × 10^-5, trained on 8 H100 GPUs. The excerpt provided does not contain explicit diffusion-loss equations, denoising objectives, or symbolic notation for dataset variables or the training objective.

5. Benchmarking, ablations, and empirical effects

TripVVT-Bench is the evaluation companion to TripVVT-10K. It is described as a benchmark “built upon the TripVVT-10K dataset” and consists of 100 high-quality and challenging triplets, fully held out from training (Shao et al., 30 Apr 2026). Its coverage includes various garment types, scenes with complex backgrounds, and both single- and multi-person videos. To support comparison with competing methods, it also provides auxiliary inputs such as garment masks and DensePose information, generated following CatV2TON’s methodology.

The benchmark defines a multi-dimensional metric suite. Video quality is measured by VFID using both I3D and ResNext backbones, plus SSIM and LPIPS when reference videos are available. Try-on fidelity is measured by CLIP-I and Gemini-SR, where Gemini-2.5-Flash automatically scores try-on success rate. Background consistency is evaluated with BG-L1-Err and BG-DINO-Err on non-garment regions. Temporal consistency is measured by CLIP-F between adjacent frames.

On TripVVT-Bench, the reported result for the TripVVT method is VFID-I 20.7245, VFID-R 0.3163, SSIM 0.8538, LPIPS 0.1053, CLIP-I 0.7110, CLIP-F 0.9606, BG-L1 0.0852, BG-DINO 0.0059, and Gemini-SR 91%. The paper contrasts this with open-source academic baselines and the commercial model Kling, and attributes the result to superior video quality, garment fidelity, temporal stability, and generalization on harder in-the-wild videos.

The ablation study isolates the effect of the dataset itself. Removing TripVVT-10K and training only with the supplementary dataset degrades performance from VFID-I 20.7245, VFID-R 0.3163, SSIM 0.8538, LPIPS 0.1053, CLIP-I 0.7110, CLIP-F 0.9606, BG-L1 0.0852, and BG-DINO 0.0059 to VFID-I 22.5105, VFID-R 0.4569, SSIM 0.7605, LPIPS 0.1182, CLIP-I 0.7133, CLIP-F 0.9483, BG-L1 0.1552, and BG-DINO 0.0059. The authors interpret this as evidence that the in-the-wild diversity of TripVVT-10K is crucial for generalization to complex outdoor scenes.

The paper also reports transfer beyond its own benchmark. On ViViD-S, the method achieves VFID-I 16.4398, VFID-R 0.3315, CLIP-I 0.7267, CLIP-F 0.9670, BG-L1 0.0624, and BG-DINO 0.0059, which the authors explicitly attribute to “the strong supervision of TripVVT-10K and its supplementary training sets.”

A user study on TripVVT-Bench further supports the evaluation setup. Across 2,500 votes from 50 participants, the percentage of first-place votes is 2.6% for ViViD, 0.2% for CatV2TON, 12.2% for MagicTryOn, 17.2% for Kling, and 67.6% for the TripVVT method.

6. Release status, limitations, and broader significance

The authors state that they publicly release the dataset and benchmark, and the project page listed in the paper is https://shaodingbao.github.io/TripVVT/ (Shao et al., 30 Apr 2026). The provided text does not specify license terms, usage restrictions, a separate repository URL, privacy policy, consent procedures, or legal governance for the web-crawled videos. In encyclopedic terms, those aspects are therefore not specified in the provided paper text.

The paper also identifies a methodological limitation tied to the same coarse-prior design that TripVVT-10K enables. Because the model relies on a human mask rather than explicit garment segmentation, it must infer which clothing regions should be replaced. As a result, non-target regions can sometimes be altered. The paper gives a concrete example in which a backpack disappears and there are minor changes to the skirt and boots during garment transfer. This suggests that the coarse human-mask prior improves robustness but does not guarantee perfectly controlled editing.

In a broader research context, TripVVT-10K matters because it changes both the supervision type and the target regime of video virtual try-on. Prior resources emphasized paired or image-level settings; TripVVT-10K formalizes temporally coherent cross-garment supervision over videos, with real target sequences, complex in-the-wild scenes, and auxiliary conditioning signals aligned to a DiT-based generation framework. A plausible implication is that the dataset’s primary contribution is not merely scale, but the coupling of triplet supervision, human-mask priors, and held-out in-the-wild benchmarking into a unified experimental substrate. Under that interpretation, TripVVT-10K functions as a foundational resource for controllable, realistic, and temporally stable video virtual try-on in unconstrained environments.