I have reimplemented Stable Diffusion 3.5 from scratch in pure PyTorch
a day ago
- #StableDiffusion
- #PyTorch
- #MachineLearning
- miniDiffusion is a PyTorch reimplementation of Stable Diffusion 3.5 with minimal dependencies.
- Designed for educational, experimenting, and hacking purposes with ~2800 lines of code.
- Main model files include dit.py, dit_components.py, and attention.py.
- Text encoders are in t5_encoder.py and clip.py, with tokenizers in tokenizer.py.
- Includes implementations of VAE, CLIP, T5 Text Encoders, and tokenizers.
- Features Multi-Modal Diffusion Transformer Model and Flow-Matching Euler Scheduler.
- Repository includes experimental features and requires more testing.
- Installation involves cloning the repo and installing dependencies.
- Checkpoints for models require a Hugging Face Token.
- Project is under MIT License for educational and experimental use.