search

Shenyi-Z / ToCa

Accelerating Diffusion Transformers with Token-wise Feature Caching
MIT License
27 stars 1 forks source link

More experimental questions of Open-Sora-Plan and Latte #7

Open ZTzxj opened 5 days ago

ZTzxj commented 5 days ago

Can this method be applied to Open-Sora-Plan and Latte as in the PAB paper, and would like to ask if there is a specific code for this part if it is to be applied

Shenyi-Z commented 5 days ago

Thank you for your attention to ToCa! This should be feasible, and the ToCa method is likely to be used on a variety of similar models. ( OpenSora-Plan should be similar to OpenSora) Recently, we also made a version of image generation on FLUX, which will be uploaded after sorting out in the near future. As we said, ToCa does token-wise calculation and allocation of computational layers, which should be feasible for most methods. As for experiments on more models, we may carry out follow-up work, but there is no relevant plan in the near future. However, we believe that it will not be too difficult to carry out experiments on OpenSora-Plan.

Shenyi-Z commented 4 days ago

Hello! Seems you have mentioned code for testing FLOPs. We will update it in the next few days and try to complete it within this November~