Does this apply to unet?

[songh11]

No description provided.

[ziyu-guo]

Hi there, thank you for your interest in this work. We're doing additional experiments to measure the effectiveness of this technique on UNet architecture. Will keep you posted.

[songh11]

Hi there, thank you for your interest in this work. We're doing additional experiments to measure the effectiveness of this technique on UNet architecture. Will keep you posted.

Thanks for your reply❤️, I tried on sdxl and there are two problems(used threshold_0.08 json).

1. The result within 20 steps is abnormal.
2. It's gonna be different from the original on 50 steps.
   And could you tell me how to get smoothcache_schedule on other steps?

[kelpabc123]

Hi, the schedules only currently work on the DiT model, as while the technique works on multiple models, you need to run an initial calibration pass to generate the schedule for each architecture (as different checkpoints have different optimal caching schedules). To generate one that works for SDXL, you will need to regenerate the schedule specifically for that architecture. Stay tuned for sample code that will enable that. If you can't wait, you can also take a stab at generating the schedule yourself with SDXL based on what we described in our paper.