ANE not activated

[pechaut78]

Updated to latest comfy, your package and..
and wether I ask for CPU & GPU or with ANE I get the SAME results in terms of iterations (around 1.8 on a max 3)

if I use asitop, I see that the ANE is never activated
and if I use CPU and ANE, it uses the CPU and is slow ..

Here is my compo: drop it into comfy

cheers

[aszc-dev]

Hello.
Judging by the dual CLIP, I'm assuming you are trying to run an SDXL model (or SSD-1B)? Unfortunately, SDXL models cannot be loaded to ANE. I'm not sure if it will ever be supported - here are some open issues in apple/ml-stable-diffusion on this topic:
apple/ml-stable-diffusion#271
apple/ml-stable-diffusion#237
apple/ml-stable-diffusion#267
For this project I relied heavily on that repo and its implementation of UNets (both 1.5 and XL). If SDXL on ANE is possible, a new implementation needs to be written.

[alexqem]

What About M3 chip?

[Sarievo]

Why Apple Why

[InformEthics]

Hello. Judging by the dual CLIP, I'm assuming you are trying to run an SDXL model (or SSD-1B)? Unfortunately, SDXL models cannot be loaded to ANE. I'm not sure if it will ever be supported - here are some open issues in apple/ml-stable-diffusion on this topic: apple/ml-stable-diffusion#271 apple/ml-stable-diffusion#237 apple/ml-stable-diffusion#267 For this project I relied heavily on that repo and its implementation of UNets (both 1.5 and XL). If SDXL on ANE is possible, a new implementation needs to be written.

it is not completely true, there is a SDXL Split-Einsum model for iOS which is working with MochiDiffusion on my MBP M2 16gb. Actually I would like to import it to ComyUI, it is incredibly fast on the non Pro / Max chips. Is that possible?

[aszc-dev]

That's a feature that would be very nice to have!