Daniel Bolya

Hmm, would it be possible to just attempt to run the function and then fall back to the polyfill if there's an error? I see you suggested a try-except over...

If you're not using xformers, ToMe should increase inference speed even with 512x512 images so there's something going wrong here. A couple of steps to help debug: 1. Are you...

@liujianzhi Hmm, if you're not using xformers, you should see a speed-up in your case. I wonder if tomesd is broken for diffusers 0.11.1? I've been using `diffusers==0.14.0`. If possible,...

> > > > Sorry. I upgrade my `diffusers==0.14.0`. And I still find that `5948.799s` (`w/o TOME`) and `5623.201s` (`r=90% TOME`) in generating 2000 images. I just tried to reproduce...

Hi @alex-bene, thanks for the detailed write-up. The experiments in the paper were performed _in the original stable diffusion repo_ (namely the [runway-ml](https://github.com/runwayml/stable-diffusion) one). I think users have consistently found...

> Regarding the pipeline with sdpa disabled, do you have any idea what might be going on and how I could fix it? Do you think you could try and...

Hi @bonlime, thanks for the suggestion! Actually this is something that I considered during the development of the original ToMe, and it's in fact what other methods do already (e.g.,...

> I'm working on the implementation, would get back if it works. Looking forward to it! > What do you think of a following idea I think that's a very...

> Hi @dbolya , thank you for the discussion; I have benefited greatly. I would like to ask why you mentioned that ToMe might not be able to accelerate when...

Thanks for the write-up! I took a look, and your results look worse than I was able to get. But, that might just be a difference in prompts, since I...