Instructions to use krea/krea-realtime-video with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use krea/krea-realtime-video with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("krea/krea-realtime-video", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Diffusion Single File
How to use krea/krea-realtime-video with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Request for the Causal ODE–Pretrained (ODE-init) Checkpoint for Academic Research
Hello Krea Team and @erwann ,
Thank you very much for releasing Krea Realtime 14B and its inference code. My name is Yueyi Liu, and I am a research intern at Meow Lab, College of AI, Tsinghua University.
We are currently exploring an alternative distribution-matching approach that differs from DMD and does not require an online teacher during the distribution-matching stage. Our method has shown promising preliminary results on Wan2.1 1.3B, and we would now like to investigate whether it can scale effectively to a 14B model.
From your technical blog, we understand that Krea Realtime 14B was first initialized through causal ODE pretraining and then trained with Self-Forcing DMD. However, we were only able to locate the final checkpoint in this repository.
May I kindly ask whether your team would consider releasing the intermediate causal ODE–pretrained checkpoint—the ODE-init checkpoint obtained before the DMD stage? Reproducing this pretraining stage independently at the 14B scale would be computationally challenging for us. If a public release is not currently feasible, we would also be extremely grateful if access could be provided solely for non-commercial academic research, under any conditions your team considers appropriate.
We completely understand if there are practical or internal constraints that prevent the checkpoint from being shared. We would be happy to provide further information about our method, preliminary results, or intended use if helpful.
We are preparing this research for a potential submission to ICLR 2027. If the checkpoint can be made available, we will clearly cite the Krea Realtime 14B project and gratefully acknowledge your team’s support.
Thank you very much for your time and consideration.
Best regards,
Yueyi Liu
Meow Lab, College of AI
Tsinghua University
Hey @liuyueyi-8 ,
Thank you for reaching out! The research member who was responsible for the project is currently out of office, but should come back in a few days. Let me ask him once he's back.
- Sangwu
Hi @RE-N-Y ,
Thank you very much for the update. I really appreciate your help. I’m very interested in this project and sincerely hope it might be possible to access the initialization weights. I would be grateful if you could check with him once he’s back. I look forward to hearing from you. Thank you again!
Yueyi Liu
Sure :)
https://huggingface.co/krea/krea-realtime-video/blob/main/causal_ode_init/krea-realtime-14b-causal-ode-init.safetensors
Here you go
Citation info can be found on our github: https://github.com/krea-ai/realtime-video
Hi @erwann ,
Thank you so much for sharing the initialization weights! I really appreciate your help and the team’s support. This will be very helpful for my research, and I’ll make sure to cite your work and mention your support in our acknowledgment. Thanks again for making these resources available!
Best,
Yueyi Liu