IFM releases PAN 2, a 29B general-purpose world model simulating actions as 20s videos
pmttyji · reddit · 2026-11-11
IFM open-sourced PAN 2, a general-purpose world model: given an observation (optional image + text) and actions described in text, it generates up to 20 seconds of video (704x1248 at 24 fps) simulating how the world changes, enabling simulative reasoning where agents test candidate plans before acting.
It follows the Generative Latent Prediction (GLP) architecture: an LLM conceptual backbone writes a structured, time-grounded world-state description (events, timed actor-bound actions, camera motion); a 29B MMDiT diffusion backbone trained from scratch on hundreds of millions of annotated videos predicts the video latents; a frozen Qwen3.5-9B text encoder provides conditioning and the Wan 2.2 VAE decodes frames. One checkpoint serves both text-to-video and image-to-video in BF16. Weights, GitHub code, and a blog post are public.
More from Multimodal
- Fable 5.2 nails metallic logo 3D animation in striking demo — FinanceYF5 · 2026-11-11
- AI-native DAW lets Claude compose full songs by driving the workstation itself — FinanceYF5 · 2026-11-11
- One prompt, a 90-second animated short: Fable 5.2 builds the Blender scene and music itself — FinanceYF5 · 2026-11-11
- Five Fable 5.2 demos: from a cycling pelican to an AI-native music workstation — FinanceYF5 · 2026-11-11
- Fable 5.2 one-shots a sleepy robot on a starship animation — ChrisGPT · 2026-11-11
- AI generates Alexander's charge at Gaugamela as pure code-rendered video — imjustnewatai · 2026-11-11