PrettyU

PrettyU allows you to train a digital avatar of your own and use it to generate various rich portrait photos.

Advantages:

• No need to set complex training parameters, easy-to-use, and quick to get started
• The generated photos have greater diversity, and no template images are required during generation
• Supports 1024x1024 photos (better performance, longer time)
• Supports low-memory GPUs (minimum support is around 10GB)

Requirements

1. Linux with Ubuntu/Centos
2. Nvidia-GPU + CUDA>=11.x + CuDNN(compatible with CUDA)
3. gcc/g++ >= 6.0
4. Connected to Internet (pip/huggingface/github)

Preparation

For basic usage

• Install stable-diffusion-webui
• Download majicMIX realistic and copy it to stable-diffusion-webui/models/Stable-diffusion/
  • Set majicmixRealistic_betterV2V25.safetensors as Stable Diffusion checkpoint.
• Install extension sd-webui-additional-networks
  • Extensions -> Install from URL (https://github.com/kohya-ss/sd-webui-additional-networks.git) -> Install
• Install extension adtailer
  • Extensions -> Install from URL (https://github.com/Bing-su/adetailer.git) -> Install

For high-resolution 1024x1024

• Install extension sd-webui-controlnet
  • Extensions -> Install from URL (https://github.com/Mikubill/sd-webui-controlnet.git) -> Install
• Download Controlnet Tile Model control_v11f1e_sd15_tile.pth to directory stable-diffusion-webui/extensions/sd-webui-controlnet/models/

Installation

1. Open "Extensions" tab.
2. Open "Install from URL" tab in the tab.
3. Enter https://github.com/sleepfin/sd-webui-prettyu.git to "URL for extension's git repository".
4. Press "Install" button.
5. It may take few minutes (tensorflow/mmcv installation is time-consuming)
6. You will see the message "Installed into stable-diffusion-webui\extensions\xxx. Use Installed tab to restart".

Usage

1. Training a Lora model

   • Click Train Lora Tab
   • Upload 20-35 photos
   • Fill in the name and gender description
   • Click Start Training botton

2. Wait until model is finish training

3. Generate photos

   • Click Generate Photos Tab
   • Click Refresh Models
   • Choose one Lora Models
   • Choose Style and Resolution
   • Click Generate

4. Generated Photos will be showed in photos, you can also click Show more photos to check more photos which is classified as low quality (There is a chance that all photos are classified as low quality)

Plans

• Support higher resolution of 1024x1024 (Controlnet-tile hyper-res)
• Support Windows
• Support generating photos based on reference images
• Support 2 people photos

Contribute

Welcome to contribute to this repo Even if you do not know any coding, if you find any amazing prompt, you can still contribute to

Q&A

GPU out of memory:

If GPU memory is less than 14GB, change settings as followed:

• Settings -> prettyu -> Set Traning mini-batch size to 1
• Settings -> prettyu -> Check Enable gradient_checkpointing to save GPU memory usage in cost of longer training time

RuntimeError: cutlassF: no kernel found to launch!

Maybe caused not supported xformers on your GPU, change settings as followed:

• Settings -> prettyu -> Uncheck Enable xformers when training lora model

ONNXRuntimeError

onnxruntime-gpu not supported your CUDA version. Reinstall onnxruntime-gpu according to THIS Document.

gcc/g++ version error

We need gcc/g++ (>=6.0.0, <11.0) to compile mmcv-full, follow THIS Document. After you upgrade your gcc/g++, you have to reinstall mmcv-full:

pip uninstall -y mmcv-full
pip install mmcv-full==1.7.1 --no-cache-dir