Abstract
Recent advances in 2D virtual try-on (VTON) have largely focused on architectural innovation, yet robust performance in diverse real-world conditions remains challenging. We argue that failure modes in VTON, such as inaccurate occlusion handling, unconstrained poses, and poor coverage of under-represented garments and appearance factors, are tightly coupled to dataset design choices and long-tail data regimes. We present a scalable data generation pipeline that coordinates specialized editing and generation models with vision-language model (VLM) guidance to synthesize triplets without requiring parsing masks or dense-pose annotations.
Using this pipeline, we create CURVTON-205K, a curated synthetic dataset of 205K triplets: source person, target garment, and corresponding try-on image. CURVTON-205K is explicitly designed to broaden coverage across garments, person appearance, poses, occlusions, and visual conditions. Additionally, CURVTON-205K is organized into three difficulty tiers, enabling controlled evaluation under varying levels of difficulty and supporting difficulty-aware experimentation. We analyze the dataset diversity and demonstrate its potential to improve try-on fidelity across four VTON architectures and its cross-dataset generalization.
Method
Garments, people and difficulty edits are generated in separate streams, merged into try-on triplets, and checked by a VLM that sends failures back for regeneration.
The feedback loop, attempt by attempt
Four successive generations per example, taken from the paper (Figs. 13 & 15). The VLM assessor checks every attempt; a rejected try-on goes back to the generator with a refined prompt, and the last attempt is the one it accepts.
- Garment faithfulness
- Placement plausibility
- Semantic consistency
- No visible artifacts
Pick an example above; each one plays through its attempts.
Across the dataset, the loop removes most artifacts
Artifact rate per refinement round
Three difficulty tiers
Each source person is edited with tier-specific instructions that raise pose deviation, occlusion, clutter and viewpoint change, giving 100K edited people per tier. The tiers show where a model breaks, not just its average score.
Comparison with existing datasets
CURVTON-205K is the only dataset here that is mask-free, in-the-wild, paired, difficulty-tiered, and diverse in both garments and people (Table 1).
Diversity coverage: Limited = neither garment nor person-appearance diversity targeted; Moderate = one; Broad = both. All listed datasets are public except WildVTON.
Results
Training on CURVTON-205K improves models on its own test set, transfers to real benchmarks, and wins in-the-wild comparisons.
CURVTON-205K test set (Tables 2 & 8). UB/LB = upper/lower body, G = full-body garments, Tr/NTr = traditional/non-traditional. Hard stays the hardest tier after training.
Train on real data, CURVTON-205K, or both; test protocols fixed (Table 3).
OOTDiffusion trained on CURVTON-205K is preferred by human raters for both try-on quality (35%) and realism (28%), and scores highest on CLIP-I and the VLM evaluator.
OOTDiffusion trained on each data source; pretrained FLUX and IDM-VTON as references (Table 5). Human preference: blind pairwise votes from 20 annotators, normalized to 100% per criterion.
Explore the data
Compare any source person with its try-on, then browse more try-ons and the garments behind them. All samples come from the public sample split on Hugging Face.
Drag the handle (or use ← →) to compare the source person with the synthesized try-on; the strip picks examples from all three difficulty tiers.
Hover a card to see the source person; clickTap a card to open the side-by-side comparison.
Hover to peek at the try-on, clickTap to compare. The garment dictionary covers 200 types, from common Western clothing to long-tail and traditional garments. Each garment here opens the try-on it was used for.
Dataset
Released on Hugging Face as six tar archives (difficulty × presentation), about 108 GB in total, plus a small sample split.
Download
from huggingface_hub import snapshot_download # sample split only (remove allow_patterns for all) path = snapshot_download( repo_id="curvton2/curvton2", repo_type="dataset", allow_patterns="sample_dataset/*", )
Layout
easy/ medium/ hard/
└── female/ male/
├── cloth_image/ # garment I_c
├── initial_person_image/ # person I_p
└── tryon_image/ # try-on I_t
Matching file names pair a triplet. Code for the pipeline, tier dictionaries, evaluation and diversity metrics: github.com/curvton/CURVTON-205K_CODE.
Citation & license
BibTeX
@inproceedings{bartaula2027curvton205k,
title = {{CURVTON-205K}: A Mask-Free
Difficulty-Stratified Synthetic
Dataset for Robust and Scalable
Virtual Try-On},
author = {Bartaula, Dipan and Shrestha, Anuj
and Belbase, Ankit and Bishwakarma,
Anish and Poudel, Saroj and Ghimire,
Samip and Bhatta, Shashank and
Arazo, Eric and Bhattarai, Binod},
booktitle = {Proceedings of the IEEE/CVF Winter Conference
on Applications of Computer Vision (WACV)},
year = {2027}
}
License
The dataset is released under Creative Commons Attribution-NonCommercial 4.0, with the additional non-commercial restrictions of the FLUX.2 [klein] 9B model used to generate it.
Intended for academic, non-commercial research: virtual try-on, garment transfer, controllable image generation, curriculum learning and benchmarking.
Not for:
- identity, face or biometric recognition, or sensitive-attribute inference;
- surveillance, medical or clinical decisions;
- deceptive or non-consensual edited imagery.
Female/male labels organize the data only; they are not identity or demographic annotations.




