Stable Diffusion optimized web UI with low vram, up to 1216x1216 on 8 gb vram

github.com

41 points·avocado2··8 comments

I'm continually stunned with both the quality and variability of modifications made to StableDiffusion in the short amount of time in which this model has been available... and I think we're just scratching the surface of what's to come.

Bra-freakin'-vo to this chap.

My laptop is old but I really want to try this. What options do I have to use this? Speaking as a total n00b can I install and run this on a EC2 instance? What is the minimum instance type (like c4.xlarge, etc) recommended for running this?

I wish there was a ready-made way to get access to the UI with all the options enabled and a fixed monthly price with hopefully a RESTful api. This is so amazing. Great work!

google colab would probably be the easiest way to get started.

Would this run on Google Colab? The free tier maxes out at about 12 gigs of vram, which is easy to hit when running the original unoptimized version.

I don't know about this UNI, but I've gotten it working in Colab.

This seems to advertise being able to run on GPUs with smaller RAM, down to 2GB - but no such luck here on my laptop 3050Ti 4GB. Does anyone have any clues on the config or setup to get this going on smaller GPUs?

pretty neat. I heard rumors that generating image this large isn't always great since it was initially trained on 512x512. Any truth to that?

Biggest I could do without a crash with the HLKY version was 640x512.

it depends on what effects you're looking to get. Going beyond 512x might produce near duplicate subjects or other repeating shapes. I exploited this behavior for a series of prompts that were all about high contrast geometric shapes. This was not so satisfactory for other subjects though.

Reply on news.ycombinator.com