CatVTON (open-weight VTON)

CatVTON (ICLR 2025) concatenates the person and garment in latent space on SD 1.5 inpainting, skipping text cross-attention. The trainable attention adapters are small (~50M); full inference is about 899M params and typically under 8GB VRAM at 1024×768.


Registry id	`catvton`
Weights	`zhengchong/CatVTON`
Code	github.com/Zheng-Chong/CatVTON
Paper	arXiv:2407.15886
VRAM	<8GB @ 1024×768 (fp16/bf16)
License	CC BY-NC-SA 4.0 (code + checkpoints) — not for commercial D2C

The same HF repo also contains a FLUX.1-Fill LoRA (flux-lora/, 37.4M). Official FLUX inference code was not released with the LoRA; OpenTryOn implements the documented SD 1.5 pipeline (mix / vitonhd / dresscode attention folders).

Install

pip install opentryon[local]

The SD 1.5 inpainting base (runwayml/stable-diffusion-inpainting) may be gated. Override with a community mirror:

export CATVTON_BASE_MODEL=botp/stable-diffusion-v1-5-inpainting

Environment

# CATVTON_ATTN_CKPT=zhengchong/CatVTON
# CATVTON_BASE_MODEL=runwayml/stable-diffusion-inpainting
# HF_TOKEN=hf_...

CLI

opentryon vton --model catvton \
  --person-image person.jpg --garment-image dress.jpg \
  --garment-type dresses --attn-version mix --dry-run

opentryon vton --model catvton \
  --person-image person.jpg --garment-image top.jpeg \
  --mask-image agnostic.png --steps 50 --width 768 --height 1024

--attn-version mix is the 1024 mix checkpoint (default). vitonhd and dresscode are the 512 dataset-specific adapters.

Pass --mask-image (white = replace) when you have an agnostic mask. Without it, OpenTryOn uses a geometric upper/lower/dress rectangle.

Python

from tryon.models import CatVTONAdapter

adapter = CatVTONAdapter(attn_version="mix")
images = adapter.generate_and_decode("person.jpg", "garment.jpg")
images[0].save("tryon.png")

MCP

Tool: vton_catvton. Needs pip install opentryon[local] + GPU. Restart MCP so TryOn Studio lists it.

Install​

Environment​

CLI​

Python​

MCP​

See also​

Install

Environment

CLI

Python

MCP

See also