What combinations of models and controlnets work together?

#16
by overthruster - opened

With Guernika 3.0.2, I deleted and redownloaded models. Downloaded ControlNets. Keep getting errors and crashes trying to generate images. The errors are: "Failed to obtain prediction for sample 0", and another I can't reproduce now. I'm trying with Inkpunk Diffusion and Human Pose 1.5 as modeled in the App Store screenshot.

Is there something I'm missing?

Hello,
I noticed this bug happening when using Neural Engine (or All Compute) mode with controlNet,
You should use GPU mode until we get a fix :)

Cheers.

Guernika org

I will take a look at this, initially unless you use a different output size, ControlNet should work with the neural engine.

Hello @GuiyeC !
I'm still experiencing the same issues with CPU+NE+ControlNet,
It probably won't help, I found something interesting on "Failed to obtain prediction for sample 0" error.
It happens when you preload controlNet before pressing generate (normal behavior).

However,

  • I was mid-generation on ANE I pressed [LOAD] on controlNet,
  • I dropped an image before the generation ended and it became slow (sudo powermetrics showed it went in CPU-only mode)
  • When the first gen ended, the next ones where (controlnet)working but CPU mode

I have M1 Max, does it mean m1 ANE's are too weak ?

Guernika org

I still have to be able to reproduce this, I have changed quite a lot of how loading works so this might be fixed but it could be that some models just don't work on CPU&NE

Sign up or log in to comment