What it does Disco Diffusion is a Google Colab notebook that generates AI art and animations from text prompts using guided diffusion models and CLIP. It started as Katherine Crowson’s original notebook and accumulated features like 2D/3D animation, video input, keyframing, depth estimation, and VR output through a long chain of contributors.

The interesting bit The project describes itself as a “frankensteinian amalgamation” — and means it. Development happened by tacking on new techniques (Dango’s cutout method, Chigozie’s keyframing, MiDaS depth estimation, optical flow warping) until the changelog reads like a patchwork oral history of 2021-2022 generative art. The README’s “[to be updated with further info soon]” has apparently been pending since at least 2022.

Key highlights

• Supports 2D and 3D animation modes with camera movement (zoom, pan, rotation)
• Video input with optical-flow-based “Warp mode” for smoother results
• VR mode output for immersive content
• Custom model support including niche fine-tuned models (Pixel Art, Watercolor, Pulp SciFi, portrait generator)
• Uses Colab-Convert to let developers edit Python files instead of wrestling with notebook JSON directly
• Symmetry controls, multiple CLIP model evaluation, and prompt keyframing for temporal variation

Caveats

• Effectively frozen since late 2022; the MiDaS v3 pin was the last substantive fix
• Requires significant VRAM for larger models (ViT-L/14@336px explicitly noted as “high VRAM”)
• The “unexposed batch option since it doesn’t work” from v1.1 suggests some features were abandoned rather than fixed

Verdict Worth studying as a time capsule of pre-Stable-Diffusion experimentation, or if you need specific legacy techniques like the depth-based 3D pipeline. Skip it if you just want to generate images today — modern tools are faster, smaller, and maintained.