I started the Stable Diffusion repository in 2026-10.
It is a PyTorch implementation of Stable Diffusion v1.5 written from scratch.
The repository implements the CLIP text encoder, VAE encoder and decoder, U‑Net, and DDPM sampler without using diffusers.
The demo runs on CPU by default.
Running a 512 × 512 text‑to‑image generation with 50 DDPM steps and CFG scale 8 takes about 5 min 22 s, roughly 6.45 s per step.