IFM open-sources PAN2: 29B latent diffusion video model, 20s 720p, training code included
Aggressive-Zebra-949 · reddit · 2026-11-11
IFM released PAN 2 on Reddit, a 29B-parameter latent diffusion video model under Apache 2.0, generating up to 20 seconds of 720p 24fps video.
- Weights on Hugging Face with Diffusers and FastVideo support; VLLM coming soon
- Unusually, full training code is also open-sourced (github.com/ifm-ai/PAN2) plus a technical blog
- The LLM acts as a "conceptual reasoner," producing dense, time-stamped text states/scripts; camera moves are nested under events and disentangled from actions, handling a dozen timed camera moves in one 20s clip
Related event: IFM Open-Sources PAN 2, a 29B-Parameter Video World Model(2 posts)→
More from Multimodal
- Hybrid AI Video Pipeline Test: MiniMax H3 for Audio, Nano Banana 2 for Visuals — RollingMeteors · 2026-11-11
- Slopus and Slopfab add MiniMax H3 Image support with editing and outpainting — BittiAI · 2026-11-11
- AI art thread: 'He ran so fast that time eventually lost track of him' — LudovicCreator · 2026-11-11
- Google Antigravity agent edits your photo, adjusting trees and figures to taste — AI_Andrew · 2026-11-11
- Watercolor diorama from one photo: Gemini nails a scroll-driven story task — AI_Andrew · 2026-11-11
- ElevenLabs ships Scribe v3: 43% fewer real-time errors, 102 languages — minchoi · 2026-11-11