Open-Source Virtual Try-On Models Compared: IDM-VTON, CatVTON, OOTDiffusion, Leffa, FASHN & 14 More, and Fashio AI

What it is. The most-used open try-on model: an SDXL-based dual network (TryonNet + GarmentNet) with an IP-Adapter for the garment, by Yisol Choi and co-authors.

• GitHub · 5,213★
• Hugging Face weights
• Demo Space · running
• arXiv 2403.05139
• ComfyUI · community, 594★
• IDM-VTON guide

Garments

Upper, lower, dresses (DressCode checkpoint)

Mask-free

No — mask + DensePose (automatic in the demo)

VRAM

Not documented

Paper

Improving Diffusion Models for Authentic Virtual Try-on in the Wild (ECCV 2024)

Training code

Yes (train_xl.py)

Code

CC BY-NC-SA 4.0

Weights

CC BY-NC-SA 4.0

Base

SDXL inpainting 0.1 + IP-Adapter

Data

VITON-HD (CC BY-NC 4.0), DressCode (non-commercial)

On Hugging Face. yisol/IDM-VTON: 745 likes, 35,306 downloads in the last 30 days, 1.61 million all time. Space: 2,137 likes, ZeroGPU.

Quick start (from the README):

git clone https://github.com/yisol/IDM-VTON.git
cd IDM-VTON
conda env create -f environment.yaml
conda activate idm
python gradio_demo/app.py

Watch out: The authors do not publish a VRAM figure or inference time. Numbers you see quoted elsewhere are community reports. The ComfyUI node is GPL-3.0, which is a licence for the node, not for the weights.

Full guide: demo walkthrough, local setup, ComfyUI install and the licence in detail are in our IDM-VTON guide.

What it is. A single Stable Diffusion 1.5 inpainting UNet that concatenates person and garment instead of adding a second network: 899M parameters in total, 49.57M trainable.

• GitHub · 1,863★
• Hugging Face weights
• Demo Space · broken
• arXiv 2407.15886
• ComfyUI · official workflow
• CatVTON guide

Garments

Upper, lower, overall / dresses

Mask-free

Mask-based (AutoMasker); a mask-free checkpoint is gated

VRAM

~8 GB at 1024×768, bf16

Paper

Concatenation Is All You Need for Virtual Try-On with Diffusion Models (ICLR 2025)

VRAM

~8 GB at 1024×768, bf16 (README)

Code

CC BY-NC-SA 4.0

Weights

CC BY-NC-SA 4.0; mask-free checkpoint gated, research-only, no sharing

Base

SD 1.5 inpainting (also a FLUX.1-Fill-dev LoRA, non-commercial base)

Data

VITON-HD / DressCode

On Hugging Face. zhengchong/CatVTON: 5,948 downloads in the last 30 days. The official Space showed RUNTIME_ERROR on 6 Oct 2026.

Quick start (from the README):

conda create -n catvton python==3.9.0
conda activate catvton
pip install -r requirements.txt
CUDA_VISIBLE_DEVICES=0 python app.py --output_dir="resource/demo/output" --mixed_precision="bf16" --allow_tf32

Watch out: The demo Space is down, so run it locally or through the official ComfyUI workflow in the repo's Releases. The mask-free checkpoint is gated and its licence forbids sharing it.

Full guide: why the demo is down, local setup, the ComfyUI workflow, CatVTON-FLUX and the licence in detail are in our CatVTON guide.

What it is. Outfitting Fusion based Latent Diffusion, from Xiao-i Research (Yuhao Xu et al.). The most-starred repo in this list at 6,604 stars.

• GitHub · 6,604★
• Hugging Face weights
• Demo Space · running
• arXiv 2403.01779
• ComfyUI · community, 471★

Garments

Upper, lower or dress — one category per run

Mask-free

Automatic (SCHP human parsing + OpenPose)

VRAM

Not documented

Paper

Outfitting Fusion based Latent Diffusion for Controllable Virtual Try-on (AAAI 2025)

Garments

upperbody (0), lowerbody (1), dress (2) — one category per run

Tested on

Ubuntu 22.04 only (README)

Code

CC BY-NC-SA 4.0

Weights

CC BY-NC-SA 4.0

Data

VITON-HD (half-body) / DressCode (full-body)

On Hugging Face. levihsu/OOTDiffusion: 339 likes. Space: 1,128 likes, running on ZeroGPU.

Quick start (from the README):

conda create -n ootd python==3.10
conda activate ootd
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
pip install -r requirements.txt
cd run && python run_ootd.py --model_path <model.jpg> --cloth_path <cloth.jpg> --scale 2.0 --sample 4

Watch out: It does one garment category per pass, not a full outfit in one shot, and the repo ships no training code. The last commit was in May 2024.

What it is. A 972M-parameter MMDiT that works in pixel space, trained from scratch by FASHN AI on their own data (18M masked pairs plus 4M synthetic triplets, per the model card).

• GitHub · 342★
• Hugging Face weights
• Demo Space · running
• ComfyUI · community, 11★

Garments

Tops, bottoms, one-pieces

Mask-free

Yes — mask-free by default

VRAM

~8 GB (model card)

VRAM

~8 GB (model card)

Speed

~5 s on an H100 (model card)

Input

Works with flat-lay garment photos

Code

Apache-2.0

Weights

Apache-2.0

Base

None (trained from scratch)

Data

FASHN's own

On Hugging Face. fashn-ai/fashn-vton-1.5: 139 likes. Space: running on ZeroGPU. The paper is listed as coming soon.

Quick start (from the README):

pip install -e .            # inside the cloned repo
python scripts/download_weights.py --weights-dir ./weights
# then, in Python:
from fashn_vton import TryOnPipeline
pipeline = TryOnPipeline(weights_dir="./weights")
result = pipeline(person_image=person, garment_image=garment, category="tops")

Watch out: The bundled human parser has its own licence. Check it separately before shipping.

What it is. Learning Flow Fields in Attention, from Zijian Zhou et al. (CVPR 2025). It does virtual try-on and pose transfer from one codebase.

• GitHub · 1,681★
• Hugging Face weights
• Demo Space · running
• arXiv 2412.08486
• ComfyUI · community, 241★

Garments

Upper, lower, dresses (DressCode checkpoint)

Mask-free

No — SCHP + DensePose (automatic in the demo)

VRAM

Not documented

Paper

Learning Flow Fields in Attention for Controllable Person Image Generation

Speed

~6 s on an A100, fp16 (model card)

VRAM

Not documented by the authors

Code

MIT

Weights

MIT (HF card)

Base

SD inpainting (try-on), SDXL inpainting (pose transfer)

Data

VITON-HD and DressCode, both non-commercial

On Hugging Face. franciszzj/Leffa: 344 likes; the try-on checkpoints are 7.2 GB each, pose transfer 20.9 GB. Space: 629 likes, running.

Quick start (from the README):

conda create -n leffa python==3.10
conda activate leffa
cd Leffa
pip install -r requirements.txt
python app.py

Watch out: MIT on the code and weights does not clear the training data. The weights learned from datasets licensed for research only, which is why we mark it grey. Get legal advice before commercial use.

What it is. Semantic correspondence on a latent diffusion model, from Jeongho Kim et al. at KAIST (CVPR 2024). Built on Paint-by-Example.

• GitHub · 1,266★
• Hugging Face weights
• Demo Space · broken
• arXiv 2312.01725

Garments

Upper body (VITON-HD)

Mask-free

No

VRAM

Not documented

Paper

Learning Semantic Correspondence with Latent Diffusion Model for Virtual Try-On

Environment

torch 2.0.0 + cu117, pytorch-lightning 1.5.0

Successor

PromptDresser

Code

CC BY-NC-SA 4.0 (README; no LICENSE file)

Weights

CC BY-NC-SA 4.0, on KAIST SharePoint

Base

Paint-by-Example

Data

VITON-HD

On Hugging Face. The official Space showed CONFIG_ERROR (“No candidate PyTorch version found for ZeroGPU”) on 6 Oct 2026.

Quick start (from the README):

git clone https://github.com/rlawjdghek/StableVITON
cd StableVITON
conda create --name StableVITON python=3.10 -y
conda activate StableVITON
# then the pinned pip installs listed in the README

Watch out: The pinned 2023 stack (torch 2.0, CUDA 11.7) is the main install hurdle. Upper body only.

What it is. A 2021 GAN (ALIAS normalisation) from Seunghwan Choi et al. The name also covers the 1024×768 benchmark dataset that most later models train on.

• GitHub · 1,169★
• arXiv 2103.16874

Garments

Upper body

Mask-free

No (GAN)

VRAM

Not documented

Paper

High-Resolution Virtual Try-On via Misalignment-Aware Normalization (CVPR 2021)

Code

CC BY-NC 4.0

Weights

CC BY-NC 4.0 (Google Drive)

Dataset

CC BY-NC 4.0, “for research purposes only”

On Hugging Face. No official Hugging Face weights or Space.

Quick start (from the README):

git clone https://github.com/shadow2496/VITON-HD.git
conda create -y -n viton python=3.8 && conda activate viton
pip install opencv-python torchgeometry
CUDA_VISIBLE_DEVICES=0 python test.py --name <NAME>

Watch out: Some sites list the code as MIT. The repo puts all material under CC BY-NC 4.0, © NeStyle.

What it is. A condition generator plus image generator GAN from Sangyun Lee et al. (ECCV 2022) that handles misalignment and occlusion.

• GitHub · 917★
• arXiv 2206.14180

Garments

Upper body

Mask-free

No (GAN)

VRAM

Not documented

Paper

High-Resolution Virtual Try-On with Misalignment and Occlusion-Handled Conditions

Training

Generator stage ~4 days on two RTX 3090s (README)

Environment

PyTorch 1.8.2 + cu111

Code

CC BY-NC 4.0

Weights

CC BY-NC 4.0

On Hugging Face. No official Hugging Face weights or Space.

Quick start (from the README):

conda create -n hrviton python=3.8
pip install opencv-python torchgeometry Pillow tqdm tensorboardX scikit-image scipy
python3 test_generator.py --occlusion --cuda True --test_name <name> --tocg_checkpoint <ckpt> --gen_checkpoint <ckpt> --datasetting unpaired --dataroot <data> --data_list <pairs.txt>

Watch out: A 2021-era PyTorch stack. Fine for a benchmark, slow to modernise.

What it is. An SD3-based diffusion transformer from Boyuan Jiang et al. at Tencent that targets garment texture and fit.

• GitHub · 632★
• Hugging Face weights
• Demo Space · running
• arXiv 2411.10499
• ComfyUI · official, 105★

Garments

Not listed in the README

Mask-free

Two steps: automatic mask you can edit, then try-on

VRAM

Not stated (offload modes)

Paper

Advancing the Authentic Garment Details for High-fidelity Virtual Try-on

Memory modes

bf16 / fp16 + offload / aggressive offload

Code

CC BY-NC-SA 4.0

Weights

CC BY-NC-SA 4.0

Base

SD3

Commercial

Offered through Tencent Cloud

On Hugging Face. BoyuanJiang/FitDiT: 214 downloads in the last 30 days. Space: 280 likes, running. The team also released the CVDD dataset.

Quick start (from the README):

pip install torch==2.4.0 torchvision==0.19.0 diffusers==0.31.0 transformers==4.39.3 gradio==5.8.0 onnxruntime-gpu==1.20.1
python gradio_sd3.py --model_path local_model_dir --fp16 --offload

Watch out: No VRAM number is published. The offload flags trade speed for memory, so start with --fp16 --offload on a smaller card.

What it is. A mask-free FLUX.1-Fill LoRA from Kunbyte AI and Zhejiang University that covers 15 wearable classes, not just clothes.

• GitHub · 273★
• Hugging Face weights
• Demo Space · broken
• arXiv 2508.13632

Garments

15 classes: tops, bottoms, dresses, shoes, earrings, bracelets, necklaces, rings, sunglasses, glasses, belts, bags, hats, ties, bow ties

Mask-free

Yes — mask-free

VRAM

≥28 GB, bf16 (README)

Paper

Virtual Try-On Anything without Masks

VRAM

≥28 GB, bf16 (README)

Classes

Tops, bottoms, dresses, shoes, earrings, bracelets, necklaces, rings, sunglasses, glasses, belts, bags, hats, ties, bow ties

Code

Apache-2.0

Weights

Apache-2.0 LoRA (344 MB)

Base

FLUX.1-Fill-dev — FLUX.1 [dev] Non-Commercial License

Data

Not stated

On Hugging Face. Kunbyte/OmniTry: 20 likes. The Space showed CONFIG_ERROR (torch version not compatible with ZeroGPU) on 6 Oct 2026.

Quick start (from the README):

conda env create -f environment.yml
conda activate omnitry
pip install -r requirements.txt
pip install flash-attn==2.6.3
python gradio_demo.py

Watch out: An Apache-2.0 LoRA is not commercial-ready if the base model it needs is non-commercial. This is the clearest example of the base-model layer below.

What it is. Not a try-on model as such. Alibaba's general image editor (Qwen-Image-Edit-2509 and 2511) plus a community LoRA trained for try-on, such as FoxBaze's Try_On_Qwen_Edit_Lora_Alpha.

• Hugging Face weights
• ComfyUI · ComfyUI workflow on the LoRA card

Garments

One person + several garment references (LoRA card)

Mask-free

Yes — instruction-based editing

VRAM

Not documented

Models

Qwen-Image-Edit-2509 · 2511 · FoxBaze try-on LoRA

ComfyUI

The LoRA card ships a workflow JSON

Base

Apache-2.0

LoRA

Apache-2.0 (FoxBaze alpha); check each LoRA separately

On Hugging Face. Qwen-Image-Edit-2509: 466,708 downloads in the last 30 days; 2511: 258,015. FoxBaze LoRA: 1,105.

Watch out: Every LoRA has its own licence and training data. “Apache-2.0 base” only clears the base.

Official demo

Kwai-Kolors/Kolors-Virtual-Try-On on Hugging Face: running on 6 Oct 2026, 10,203 likes. Also mirrored as a ModelScope studio

Official product

Virtual Try-On inside Kling AI. Kling's guide quotes 10–13 seconds per result and five output aspect ratios (1:1, 4:3, 3:4, 3:2, 2:3)

Weights

None. The Kwai-Kolors Hugging Face organisation publishes ten models (Kolors, its ControlNets, IP-Adapters, inpainting and others) and none of them is the try-on model

How the Space works

It runs on CPU. Its app.py sends your two images to a remote Kuaishou endpoint and polls for the result, so the model never runs on Hugging Face

Licence tag

The Space says Apache-2.0, but that covers the small Gradio client, not the model behind it

Kolors base model

A separate text-to-image model. Code Apache-2.0; weights open for academic research, and commercial use requires registering with Kuaishou through a questionnaire (README)

Run locally / ComfyUI

Not possible: there is nothing to download