CURVTON-205K: A Mask-Free Difficulty-Stratified Synthetic Dataset for Robust and Scalable Virtual Try-On
Organizations: IOE Pulchowk Campus, Tribhuvan University, Nepal · NAAMII, Nepal · Fogsphere (Redev.AI Ltd, UK) · University of Aberdeen, UK
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.
Figures & tables
| Dataset | Train | Test | Type | Mask- Free | Diff. tiers | In-the- Wild | Cloth Pairs | Diversity coverage |
|---|---|---|---|---|---|---|---|---|
| VITON [ 23 ] | 14K | 2K | Real | Limited | ||||
| VITON-HD | 12.3K | 1.4K | Real | Limited | ||||
| DressCode | 50K | 3K | Real | Limited | ||||
| StreetTryOn | 12.5K | 1.9K | Real | Moderate | ||||
| WildVTON | N/A | 1.2K | Hybrid | Limited | ||||
| OpenVTON-Bench | 90K | 10K | Real | Moderate |
| Overall | Difficulty Tier | Cloth Type | ||||||||
| Easy | Med. | Hard | UB | LB | G | Tr | NTr | |||
| OOTDiffusion [ 48 ] | Zero Shot | 18.90 | 25.20 | 24.40 | 25.50 | 19.50 | 27.15 | 29.70 | 29.95 | 17.15 |
| Training | 8.75 | 9.88 | 11.54 | 14.12 | 11.42 | 11.83 | 13.79 | 14.22 | 13.83 | |
| CatVTON [ 13 ] | Zero Shot | 23.85 | 22.55 | 24.40 | 27.55 | 22.60 | 23.55 | 28.90 | 31.35 | 20.16 |
| Training | 9.14 | 9.14 | 9.23 | 12.88 | 10.62 | 11.34 | 14.94 | 15.57 | 12.70 | |
| IDM-VTON [ 12 ] | Zero Shot | 17.05 | 16.20 | 17.30 | 19.30 | 16.70 | 17.25 | 22.00 | 25.55 | 15.15 |
| Model | Train Data | VITON-HD | DressCode | StreetTryOn | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| SSIM | LPIPS | FID | KID | SSIM | LPIPS | FID | KID | FID | KID | ||
| CatVTON [ 13 ] | Real baseline | 0.87 | 0.11 | 9.01 | 1.09 | 0.89 | 0.10 | 6.14 | 1.40 | 25.31 | 6.95 |
| CURVTON-205K | 0.78 | 0.13 | 11.74 | 1.41 | 0.80 | 0.12 | 7.98 | 1.79 | 21.36 | 5.42 | |
| Real + CURVTON | 0.89 | 0.10 | 8.72 | 1.04 | 0.90 | 0.10 | 5.92 | 1.36 | 20.45 | 5.18 | |
| IDM-VTON [ 12 ] | Real baseline | 0.85 | 0.12 | 9.26 | 1.27 | 0.89 | 0.12 | 14.13 | 6.03 | 23.62 | 6.18 |
| CURVTON-205K | 0.80 | 0.13 | 11.02 | 1.46 | 0.83 | 0.13 | 16.94 | 6.88 | 19.08 | 4.96 | |
| Dataset | Pose | Occlusion | Background | Illumination | Body Shape | Garment | Camera | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Occl. Cmplx. | Person Vis. | BGE (C) | BSE (D) | BOD (D) | IGM | LVM (D) | Gmt Var. | Gmt Div. | |||||
| VITON-HD | 145.55 | 0.52 | 0.31 | 2.93 | 0.01 | 0.29 | 0.60 | 0.06 | 0.54 | 475.81 | 0.04 | 0.070 | 0.090 |
| DressCode | 127.79 | 0.45 | 0.45 | 1.60 | 0.07 | 0.56 | 0.73 | 0.10 | 0.76 | 669.59 | 0.26 | 0.140 | 0.070 |
| StreetTryOn | 204.48 | 0.62 | 0.23 | 4.57 | 0.54 | 3.15 | 0.45 | 0.10 | 0.47 | 621.65 | 0.15 | 0.120 | 0.140 |
| Dress-ED | 129.92 | 0.43 | 0.43 | 1.63 | 0.07 | 0.60 | 0.80 | 0.09 | 0.77 | 725.11 | 0.32 | 0.170 | 0.113 |
| CURVTON Easy | 66.68 | 0.53 | 0.33 | 4.00 | 0.43 | 3.45 | 0.50 | 0.16 | 0.62 | 831.32 | 0.52 | 0.022 | 0.066 |
| Train Data / Model | CLIP-I | VLM | Human Preference (%) | |
|---|---|---|---|---|
| Quality | Realism | |||
| VITON-HD | 0.71 | 0.69 | 10% | 11% |
| DressCode | 0.74 | 0.73 | 12% | 13% |
| VHD + DC | 0.77 | 0.76 | 14% | 15% |
| FLUX.2 [klein] 9B (Pretrained) | 0.76 | 0.74 | 14% | 14% |
| IDM-VTON | 0.80 | 0.78 | 15% | 19% |
Appendix figures & tables31 assets
Supplementary material from the paper’s appendix.
Appendix
| Stage | Iteration (VLM Thr = 0.8) | Train Artifacts (%) | Test Artifacts (%) | ||||
|---|---|---|---|---|---|---|---|
| Easy | Medium | Hard | Easy | Medium | Hard | ||
| Person & cloth generation | 17 | 25 | 35 | ||||
| Person/Garment VLM Filtering | 3 | 4.4 | 5.7 | ||||
| Try-On Generation | Iteration 0 | 38 | 48 | 60 | 39 | 49 | 61 |
| Iteration I | 25 | 32 | 40 | 26 | 33 | 41 | |
| Iteration II | 15 | 19 | 24 | 16 | 20 | 25 | |
| Generator | Est. Cost | Generator Image Cond | Evaluator(s) | Evaluator Image Cond | Artifact Rate (%) Easy / Medium / Hard Iteration 0 | Artifact Rate (%) Easy / Medium / Hard Iteration IV |
| FLUX.2-klein-9B | $25k | 1–4 | Qwen3-VL 32B | 3 | 44 / 54 / 66 | 8 / 11 / 14 |
| Qwen3-VL-30B-A3B Thinking | 3 | 44 / 54 / 66 | 9 / 12 / 13 | |||
| PaliGemma | 1 | 44 / 54 / 66 | 13 / 16 / 20 | |||
| FLUX.2-klein-4B | $18k | 1–4 | Qwen3-VL 32B | 3 | 55 / 63 / 74 | 15 / 18 / 22 |
| PaliGemma (optional) | 1 | 55 / 63 / 74 | 18 / 22 / 27 | |||
| Qwen-Image-Edit- 2509 | $90k | 1–3 | Qwen3-VL 32B | 3 | 65 / 73 / 82 | 21 / 25 / 30 |
| Overall | Difficulty Tier | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SSIM | LPIPS | FID | LPIPS | SSIM | FID | ||||||||
| E | M | H | E | M | H | E | M | H | |||||
| OOTDiffusion [ 48 ] | Zero Shot | 0.75 | 0.19 | 18.90 | 0.18 | 0.18 | 0.22 | 0.78 | 0.77 | 0.70 | 25.20 | 24.40 | 25.50 |
| CURVTON Training | 0.85 | 0.13 | 8.75 | 0.12 | 0.12 | 0.15 | 0.86 | 0.85 | 0.83 | 9.88 | 11.54 | 14.12 | |
| CatVTON [ 13 ] | Zero Shot | 0.45 | 0.36 | 23.85 | 0.30 | 0.36 | 0.41 | 0.55 | 0.41 | 0.37 | 22.55 | 24.40 | 27.55 |
| CURVTON Training | 0.86 | 0.12 | 9.14 | 0.12 | 0.09 | 0.12 | 0.86 | 0.86 | 0.81 | 9.14 | 9.23 | 12.88 | |
| Easy Test | Medium Test | Hard Test | All Mixed | |||||||||||||
| Train Data | SSIM | LPIPS | FID | KID | SSIM | LPIPS | FID | KID | SSIM | LPIPS | FID | KID | SSIM | LPIPS | FID | KID |
| Easy only | 0.86 | 0.11 | 9.86 | 0.95 | 0.83 | 0.12 | 12.70 | 0.92 | 0.76 | 0.18 | 19.87 | 2.29 | 0.85 | 0.11 | 10.29 | 0.94 |
| Medium only | 0.82 | 0.18 | 12.89 | 1.75 | 0.83 | 0.12 | 11.02 | 0.46 | 0.78 | 0.15 | 15.32 | 0.61 | 0.82 | 0.17 | 12.52 | 1.73 |
| Hard only | 0.83 | 0.16 | 13.08 | 2.51 | 0.84 | 0.12 | 11.47 | 1.11 | 0.77 | 0.15 | 15.10 | 1.08 | 0.83 | 0.15 | 13.15 | 2.75 |
| Medium + Hard | 0.81 | 0.20 | 12.75 | 1.82 | 0.82 | 0.13 | 11.11 | 0.39 | 0.77 | 0.16 | 15.25 | 0.59 | 0.80 | 0.16 | 13.92 | 0.64 |
| Easy + Hard (full) | 0.83 | 0.16 | 11.46 | 1.20 | 0.84 | 0.11 | 10.42 | 0.48 | 0.75 | 0.17 | 15.50 | 0.95 | 0.81 | 0.15 | 13.54 | 0.70 |
| Easy | Medium | Hard | All Mixed | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Regime | Budget | Curriculum | SSIM | LPIPS | FID | SSIM | LPIPS | FID | SSIM | LPIPS | FID | SSIM | LPIPS | FID |
| Finetuning (Stable Diffusion) | Minimal (3.6k iters) | No Curriculum | 0.78 | 0.17 | 21.4 | 0.74 | 0.21 | 26.7 | 0.66 | 0.27 | 36.3 | 0.73 | 0.22 | 28.1 |
| Standard Curriculum | 0.79 | 0.16 | 20.2 | 0.77 | 0.19 | 25.3 | 0.68 | 0.25 | 33.8 | 0.75 | 0.21 | 26.4 | ||
| Limited (7.2k iters) | No Curriculum | 0.81 | 0.14 | 19.3 | 0.77 | 0.18 | 23.8 | 0.70 | 0.23 | 32.7 | 0.76 | 0.19 | 25.2 | |
| Standard Curriculum | 0.82 | 0.13 | 18.4 | 0.78 | 0.17 | 22.4 | 0.72 | 0.21 | 29.8 | 0.77 | 0.17 | 23.5 | ||
| Medium (14.4k iters) | No Curriculum | 0.83 | 0.13 | 17.9 | 0.79 | 0.16 | 21.7 | 0.73 | 0.21 | 28.4 | 0.78 | 0.17 | 22.7 | |
Virtual Try-On
Virtual try-on generates images or videos showing people wearing selected garments while preserving appearance, fit, pose, and identity.
See all 58 papers