ComfyUI support?

#30
by akiokawahito - opened

Any progress on ComfyUI support?

@JeriffCheng it should work normally, no? This is the same exact architecture as SDXL and works the same. Whatever your workflow is for SDXL should work for comfy ui.

Only things that should not work ip adapters, control nets, and sdxl loras since playground is trained from scratch completely. It is possible to create new Loras, and the other things I mentioned but you would have to train them for playground.

May I have the hidden setting parameters of this demo?
cfg, steps, sampler, etc.
thank you.
image.png

@JeriffCheng they give the basic settings and I'll summarize them.

I think the cfg of 3 is usually optimal and gives good results most of the time. cfg is same as guidance scale btw.
This model works well at 20+ steps. More is better but takes more time. I think 25-35 is a sweet spot.
The demo uses the DPM++ 2M Karras scheduler and is probably the best one for this model.
This model works well enough without negative prompts but you can add them to remove objects you don't like.

Thank you so much!

@JeriffCheng it should work normally, no? This is the same exact architecture as SDXL and works the same. Whatever your workflow is for SDXL should work for comfy ui.

Only things that should not work ip adapters, control nets, and sdxl loras since playground is trained from scratch completely. It is possible to create new Loras, and the other things I mentioned but you would have to train them for playground.

Is there currently an off-the-shelf controlnet model adapted to the playground?

@hiSandog no, I sadly don't think so.

Controlnet requires a lot of work to develop. Currently, only SD1.5 and SDXL have, as far as I know..

Controlnet requires a lot of work to develop. Currently, only SD1.5 and SDXL have, as far as I know..

thank you

Sign up or log in to comment