IDM-VTON: Improving Diffusion Models for Authentic Virtual Try-on in the Wild

This is the unofficial train code of IDM-VTON "Improving Diffusion Models for Authentic Virtual Try-on in the Wild". Most code from [IDM-VTON] https://github.com/yisol/IDM-VTON, only realize the train code for viton dataset

train result

Requirements

git clone https://github.com/luxiaolili/IDM-VTON-train/tree/main
cd IDM-VTON

conda env create -f environment.yaml
conda activate idm

Data preparation

get the cloth mask and agnostic-v3.3 and put the image_mask and agnostic-v3.3 to train and test folder

python get_mask.py ../zalando-hd-resized/train/image  image_mask agnostic-v3.3

mv image_mask ../zalando-hd-resized/train/
mv agnostic-v3.3 ../zalando-hd-resized/train/

python get_mask.py ../zalando-hd-resized/test/image   image_mask agnostic-v3.3
mv image_mask ../zalando-hd-resized/test/
mv agnostic-v3.3 ../zalando-hd-resized/test/

Chang the config

You can download idm model from https://huggingface.co/yisol/IDM-VTON/ for denspose, humanparsering, openpose, image_encoder, text_encoder

You can download vae model from https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

You need download sdxl-1.0 model from https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

download sdxl-1.0-inpainting model from https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1

stable-diffusion-xl-1.0-inpainting-0.1 unet config update "encoder_hid_dim_type":"ip_image_proj", "encoder_hid_dim":1280

stabilityai/stable-diffusion-xl-base-1.0 unet config delete "addition_embed_type":"text_time"

VITON-HD

You can download VITON-HD dataset from VITON-HD.

After download VITON-HD dataset, move vitonhd_test_tagged.json into the test folder.

Structure of the Dataset directory should be as follows.


train
|-- image
|-- image-densepose
|-- agnostic-v3.3
|-- cloth
|-- img_mask

Train

sh train.sh

or Run the following command

CUDA_VISIBLE_DEVICES=1  accelerate launch --num_processes 1 --mixed_precision "fp16" train.py \
  --pretrained_model_name_or_path="idm" \
  --inpainting_model_path="stable-diffusion-xl-1.0-inpainting-0.1" \
  --garmnet_model_path="stabilityai/stable-diffusion-xl-base-1.0" \
  --width=768 \
  --height=1024 \
  --data_json_file="vitonhd.json" \
  --data_root_path="../zalando-hd-resized" \
  --mixed_precision="fp16" \
  --train_batch_size=1 \
  --dataloader_num_workers=6 \
  --learning_rate=1e-05 \
  --weight_decay=0.01 \
  --output_dir="idm_plus_output_up"\
  --save_steps=50000

Test

put the train model in the idm folder and simply run with the script file.

cd checkpoint-50000 
mv pytorch_model.bin diffusion_pytorch_model.bin
mkdir unet
mv * unet
cp ../stable-diffusion-xl-1.0-inpainting-0.1/unet/config.json unet
sh tesh.sh

or Run the following command:

CUDA_VISIBLE_DEVICES=0 accelerate launch --mixed_precision "fp16" test.py \
--pretrained_model_name_or_path="idm"

python gradio_demo/app.py

Acknowledgements

Thanks [IDM-VTION] https://github.com/yisol/IDM-VTON for most codes.

Thanks ZeroGPU for providing free GPU.

Thanks IP-Adapter for base codes.

Thanks OOTDiffusion and DCI-VTON for masking generation.

Thanks SCHP for human segmentation.

Thanks Densepose for human densepose.

License

The codes and checkpoints in this repository are under the CC BY-NC-SA 4.0 license.