Text-to-Image
Diffusers
Safetensors

Method for distillation

#2
by diuibyang - opened

Hello,
Thanks for your great work. Just wonder if you have plan to open source the method for distillation? I also noticed that you mentioned the acceleration for the z-image model, do you also have plan to release the distilled version?

inclusionAI org
β€’
edited 2 days ago

Hi, thanks for attention on TwinFlow! We are still organizing our code on large models, we will open source a version on small models then.

We are working on Z-Image, once we finished, we will also release it.

Amazing work! 1-step Z-Image is an exciting proposition. By my calculations, Z-Image needs to converge in ~2 steps to put it on equal footing with SDXL DMD2 (8 steps) in terms of speed. πŸ™‚

inclusionAI org

Amazing work! 1-step Z-Image is an exciting proposition. By my calculations, Z-Image needs to converge in ~2 steps to put it on equal footing with SDXL DMD2 (8 steps) in terms of speed. πŸ™‚

Thanks for the information, we are working on Z-Image, stay tuned!

Can this have gguf version??

Amazing work! 1-step Z-Image is an exciting proposition. By my calculations, Z-Image needs to converge in ~2 steps to put it on equal footing with SDXL DMD2 (8 steps) in terms of speed. πŸ™‚

Thanks for the information, we are working on Z-Image, stay tuned!

Can we also get distilled version of Kandinsky 19B, it seems like everyone focuses on image models, video models need expnentially more compute so Kandinsky 5 Pro distillation will be so good because it will bring down the computation requirements by a large margin compared to image models. https://huggingface.co/kandinskylab/Kandinsky-5.0-T2V-Pro-sft-10s

Sign up or log in to comment