FitVTON: Fit-aware Virtual Try-On via Body-Garment Size Control

FitVTON teaser: compared with prompt-driven commercial editing, FitVTON adapts hem position, tightness, and cuff elasticity to each body shape — With garment-body size prompts, commercial editing models produce a “neutral fit” across different body shapes. FitVTON produces faithful fit: hem position, tightness, and cuff elasticity vary across bodies.

Abstract

FitVTON is a fit-aware virtual try-on model that generates authentic garment fitting effects across diverse body shapes. FitVTON encodes garment-body size as structured text prompts (e.g., “long-length upper garment” on a “slim, medium-tall body”) and learns fitting dynamics from physically simulated try-on triplets. A FLUX.1 Kontext flow-matching backbone is fine-tuned with modality-specific LoRA adapters and dual-branch garment/body mask supervision, then rectified on real images in a second stage to bridge the sim-to-real gap.

GarmentCodeVTON Explorer

Drag either orbit to rotate · click a card to bring it to front · use the arrows to switch body. Top orbit: garments · center: SMPL-X body · bottom orbit: simulated try-on for the current body.

Body 1 / 10

Method

FitVTON architecture: FLUX.1 Kontext backbone with modality-specific LoRA adapters, dual-branch mask supervision, and two-stage training — **Top:** Person + garment + Garment-Body Size prompt → fit-aware try-on via FLUX.1 Kontext with dual LoRA adapters and dual-branch mask supervision. **Bottom:** Stage I fits on simulation triplets; Stage II rectifies textures on real images with the image LoRA only.

Highlights

Simulation data pipeline. GarmentCodeVTON: 78K aligned triplets from GarmentCode + Warp XPBD (19 garments × 16 bodies × 10 poses; one-piece / tucked-in / untucked).
Fit-aware flow-matching. FLUX.1 Kontext with dual LoRA adapters: text controls fit geometry, image handles garment transfer.
Dual-branch mask supervision. Training-only U-Net heads on garment/body masks; mask-free at inference.
Texture rectification. Stage II updates image LoRA on VITON-HD / DressCode pseudo-triplets; text LoRA frozen.
FittingEffect3K. Real-world benchmark (3,350 triplets) with VLM fit scoring aligned with human preference.

Results

Fit-oriented protocol on FittingEffect3K (GPT-scored, 1–5; category averages across GB / T-L / SC / LF)

Method	Upper Avg	Lower Avg	Dress Avg	Whole Avg
CatVTON	2.62	2.09	1.95	2.30
OmniTry	3.00	2.15	2.40	2.55
Any2AnyTryOn	2.92	2.47	1.79	2.57
JCo-MVTON	2.96	2.71	2.15	2.74
Nano Banana	3.19	2.45	2.83	2.82
FitVTON	3.22	2.99	2.90	3.08

Human preference study on FittingEffect3K (20 participants, 100 cases, 2,000 selections; best fit vs. ground truth)

Method	Selections ↑	Ratio ↑
FitVTON	666	33.30%
Nano Banana	517	25.85%
JCo-MVTON	421	21.05%
OmniTry	163	8.15%
Any2AnyTryOn	147	7.35%
CatVTON	86	4.30%

Resources

Paper: arXiv:2606.12012
Code: github.com/ZenoNing/FitVTON
Pretrained weights: huggingface.co/ZenoNing/FitVTON
GarmentCodeVTON dataset (78K simulation triplets): huggingface.co/datasets/ZenoNing/GarmentCodeVTONDataset
FittingEffect3K benchmark (3,350 real-world triplets): huggingface.co/datasets/ZenoNing/FittingEffectDataset

BibTeX

@article{ning2026fitvton,
  title   = {FitVTON: Fit-aware Virtual Try-On via Body-Garment Size Control},
  author  = {Ning, Yiqun and Shen, Ao and He, Chenhang and Zhang, Lei},
  journal = {arXiv preprint arXiv:2606.12012},
  year    = {2026}
}