Instructions to use IFM/PAN2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use IFM/PAN2 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("IFM/PAN2", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Question
Regarding open-source, open-weight, and cloned AI video and image generation models: currently, among closed-source models, the Seedance models lead in video generation, while for image generation, models like GPT-2.6, Nano Banana Pro, and 2.1 dominate the market. Now, turning to companies like yours—why do you waste money training such "garbage" models only to release them to the public? No open-source model has yet reached the level of closed-source counterparts. If closed-source models are already leading, what is the point of training models that fail to meet any real needs? For instance, the SD2.5 model produces mind-blowingly realistic video, whereas existing open-source video models—such as Minimax H3, LTX 2.5, and your own—fall far short. What is the actual purpose here? Please, conduct a thorough analysis of the AI market before training any models. Instead of just releasing open-source models for free, aim to develop one that matches the high-quality, sophisticated results of closed-source models. Do not waste capital on models that serve no practical purpose; frankly, the whole endeavor is quite ridiculous and foolish.
Hi darkenHUB, we noticed you've posted the same comment on several other new models. Fitting for the name, though we hope you're not out to darken the whole open community 🙂 Let me clarify where PAN 2 stands.
PAN 2 is a world model, not a video generator. It isn't built for the prettiest clip. It predicts what happens next when an action is taken, so an agent can test a plan before acting on it. A flawless-looking video can still get the physics wrong. For cinematic output, Seedance is excellent.
We focus on six dimensions of physical intelligence: physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamism. There, PAN 2 is ahead of the model you hold up as the standard:
- Physics-IQ Verified: PAN 2 49.4, Seedance 2.5 42.4
- RBench: PAN 2 57.0, Seedance 2.5 56.4
- Multi-step plans completed in full: PAN 2 34.0%, Seedance 2.5 24.5%
We release the weights because a closed API can't be inspected, reproduced or built on, and open research needs all three.
Different goals, different yardsticks. I believe both kinds of model have their place.
Check full details in the PAN 2 tech blog: https://ifm.ai/blog/pan2/
Ah, I see it’s just that your position in the market is different
@gxchris95 You are amazing! Any plans for an audio version? If it will have audio, it will even be comparable with H3
Watching your progress with interest now
LTX-Video are done for, that's what I'm sure, after ruining their model and reputation, you are one of the few hopes left to sustain the competition
Actual clown

