PAN 2 Released
retained text
Open weights (29B latent diffusion model): [https://huggingface.co/IFM/PAN2](https://huggingface.co/IFM/PAN2) Diffusers: [https://huggingface.co/docs/diffusers/main/en/api/pipelines/pan2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pan2) FastVideo: [https://haoailab.com/FastVideo/cookbook/pan/pan2/](https://haoailab.com/FastVideo/cookbook/pan/pan2/) VLLM: [https://recipes.vllm.ai/IFM/PAN2](https://recipes.vllm.ai/IFM/PAN2) Up to 20 seconds, 720p, 24fps Apache 2.0 License Something I think is pretty neat, we release the TRAINING code: [https://github.com/ifm-ai/PAN2/](https://github.com/ifm-ai/PAN2/) Also technical details if you're curious: [https://ifm.ai/blog/pan2/](https://ifm.ai/blog/pan2/) For our model, we give a lot of responsibility to the LLM (which we think of more as a conceptual reasoner, a lot more control than a prompt rewriter) to generate these super dense text-based state/scripts. Events and actions are timed, and camera is nested under event but disentangled from action. During training, I would often test it by trying to generate a 20 second video with like a dozen or more (timed) camera moves and it does pretty well. Feedback very welcome. I read the rules carefully, mods please let me know if there are any issues. No payment involved with this at all, so I thought it was okay to post.
document fetched 2026-11-11 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xt0yur/pan_2_released