OOTDiffusion

A packaged version of OOTDiffusion that works with Pip.

No need to manually download models, checkpoints, weights etc. Should work out if the box.

Needs CUDA and GPU.

Try on Replicate: dc (full body)

Try on Replicate: hd (half body)

Instructions

pip install git+https://github.com/qiweiii/oot_diffusion.git

Examples for Colab (T4 GPU is enough). But can be used anywhere.

If you don't set hg_root, a folder called ootd_models will be created in your working dir.

Load model

from oot_diffusion import OOTDiffusionModel
from PIL import Image
from pathlib import Path


def get_ootd_model():
  model = OOTDiffusionModel(
    hg_root="/content/models",
    cache_dir="/content/drive/MyDrive/hf_cache",
  )
  return model

Generate image

def generate_image():
  model = get_ootd_model()
  generated_images, mask_image = model.generate(
      model_path="/YOUR_MODEL.jpg",
      cloth_path="/YOUR_GARMENT.jpg",
      steps=10,
      cfg=2.0,
      num_samples=2,
    )

  return generated_images, mask_image

Display images

from IPython.display import display

generated_images, mask_image = generate_image()

for image in generated_images:
  display(image)

display(mask_image)

Credits

The original author of packaged ootd

The original author of oot cog samples

The original authors of OOTDiffusion

The authors of ComfyUI-OOTDiffusion, who made it easier to package the code.

Official implementation

See oms-Diffusion for the official implementation of OOTDiffusion.

Notes

This repo is created on the shoulder of amazing projects.

I created this repo for deploying the full-body version to replicate.

[cog-dc] for full-body api
[cog-hd] for upper body