HCP-Diffusion V2

📘中文说明

📘English document 📘中文文档

Old HCP-Diffusion V1 at main branch

Introduction

HCP-Diffusion is a Diffusion model toolbox built on top of the 🐱 RainbowNeko Engine.
It features a clean code structure and a flexible Python-based configuration file, making it easier to conduct and manage complex experiments. It includes a wide variety of training components, and compared to existing frameworks, it's more extensible, flexible, and user-friendly.

HCP-Diffusion allows you to use a single .py config file to unify training workflows across popular methods and model architectures, including Prompt-tuning (Textual Inversion), DreamArtist, Fine-tuning, DreamBooth, LoRA, ControlNet, ....
Different techniques can also be freely combined.

This framework also implements DreamArtist++, an upgraded version of DreamArtist based on LoRA. It enables high generalization and controllability with just a single image for training.
Compared to the original DreamArtist, it offers better stability, image quality, controllability, and faster training.

Installation

Install pytorch

Install via pip:

pip install hcpdiff
# Initialize configuration
hcpinit

Install from source:

git clone https://github.com/7eu7d7/HCP-Diffusion.git
cd HCP-Diffusion
pip install -e .
# Initialize configuration
hcpinit

Use xFormers to reduce memory usage and accelerate training:

# Choose the appropriate xformers version for your PyTorch version
pip install xformers==?

🚀 Python Configuration Files

RainbowNeko Engine supports configuration files written in a Python-like syntax. This allows users to call functions and classes directly within the configuration file, with function parameters inheritable from parent configuration files. The framework automatically handles the formatting of these configuration files.

For example, consider the following configuration file:

dict(
    layer=Linear(in_features=4, out_features=4)
)

During parsing, this will be automatically compiled into:

dict(
    layer=dict(_target_=Linear, in_features=4, out_features=4)
)

After parsing, the framework will instantiate the components accordingly. This means users can write configuration files using familiar Python syntax.

✨ Features

Getting Started

Training

HCP-Diffusion provides training scripts based on 🤗 Accelerate.

# Multi-GPU training, configure GPUs in cfgs/launcher/multi.yaml
hcp_train --cfg cfgs/train/py/your_config.py

# Single-GPU training, configure GPU in cfgs/launcher/single.yaml
hcp_train_1gpu --cfg cfgs/train/py/your_config.py

You can also override config items via command line:

# Override base model path
hcp_train --cfg cfgs/train/py/your_config.py model.wrapper.models.ckpt_path=pretrained_model_path

Image Generation

Use the workflow defined in the Python config to generate images:

hcp_run --cfg cfgs/workflow/text2img.py

Or override parameters via command line:

hcp_run --cfg cfgs/workflow/text2img_cli.py \
    pretrained_model=pretrained_model_path \
    prompt='positive_prompt' \
    negative_prompt='negative_prompt' \
    seed=42

📚 Tutorials

• 🧠 Model Training Guide
• 🔧 LoRA Training Tutorial
• 🎨 Image Generation Guide
• ⚙️ Configuration File Explanation
• 🧩 Model Format Explanation

Contributing

We welcome contributions to support more models and features.

Team

Maintained by HCP-Lab at Sun Yat-sen University.

Citation

@article{DBLP:journals/corr/abs-2211-11337,
  author    = {Ziyi Dong and
               Pengxu Wei and
               Liang Lin},
  title     = {DreamArtist: Towards Controllable One-Shot Text-to-Image Generation
               via Positive-Negative Prompt-Tuning},
  journal   = {CoRR},
  volume    = {abs/2211.11337},
  year      = {2022},
  doi       = {10.48550/arXiv.2211.11337},
  eprinttype = {arXiv},
  eprint    = {2211.11337},
}