hkproj/pytorch-stable-diffusion

Implements the complete diffusion pipeline including text tokenization (via BPE vocabulary), VAE-based latent encoding/decoding, and cross-attention conditioning between text embeddings and image generation. Supports loading pretrained checkpoints from Hugging Face, including fine-tuned models like InkPunk and Illustration Diffusion variants up to v1.5. Designed as an educational reference implementation that faithfully reproduces the original architecture while remaining compatible with the standard model weights ecosystem.

1,037 stars. No commits in the last 6 months.