Has this method been attempted on CogVideoX-2B?

[songh11]

I have calculated the L1 error according to the formula provided in the paper, but the errors are around 0.18. I would like to know if this method supports CogVideoX-2B. Could you please provide any insights or guidance on this? Thank you.

[kelpabc123]

Hi, sorry for the late reply. You should try different alpha values, as different architectures will require different alpha values. In theory this should work, but as this is an adjustable parameter, you should experiment with it to find a value that works for your use case! We will also have an official implementation to compute these curves and the corresponding schedule soon. Stay tuned!

[songh11]

Hi, sorry for the late reply. You should try different alpha values, as different architectures will require different alpha values. In theory this should work, but as this is an adjustable parameter, you should experiment with it to find a value that works for your use case! We will also have an official implementation to compute these curves and the corresponding schedule soon. Stay tuned!

Thank you for your response. This method is applicable to the cogvideox-2B model. Additionally, I believe this method can be used together with teacache. However, there might be overlapping acceleration when using them together. Do you have any suggestions on using these two methods concurrently?❤️

[kelpabc123]

Oops, sorry for the late reply. Thanks for pointing out TeaCache. They are very similar in implementation, although they only look at the final output rather than the outputs of individual components. If you want to stack both of them together, I recommend applying SmoothCache first then TeaCache. Let me know how it goes!

[ziyu-guo]

Following up here, we recently release v0.1 where we added the tooling to generate caching schedules yourself. We also added an example to showcase how to do it for the Diffusers DiT pipeline. Let us know if this is helpful in your experiments.