Nehno

I built a new desktop a little over a year ago after my last one's motherboard died after 8 years. During which, I updated my GPU to a RTX 4080 Super from a GTX 970 since I knew I wanted to play around with AI.

This blog was very helpful in getting Stable Diffusion set up:
https://stable-diffusion-art.com/install-windows/#RTX_4000_series_graphic_cards

I had downloaded a few different models to try and the Automatic1111 Webui seemed to be a good starting point for beginners trying out Stable Diffusion. Getting it to run was no issue at all. However, when I first tried to generate simple and small images it would take half an hour which didn't make any sense as I had a 4080S. After a few hours of trying to figure out what were the correct settings to use, I finally landed on this

set COMMANDLINE_ARGS=--xformers --opt-sdp-no-mem-attention --no-half-vae --opt-channelslast --disable-model-loading-ram-optimization

Now when using the sd_xl_base_1.0safetensors model, 512 x 512 images generated in seconds and larger 1900 x 1200 images would take up to 2 minutes. A significantly better improvement over 30 minutes for a 128 x 128 image.

This blog was very helpful in getting Stable Diffusion set up:
https://stable-diffusion-art.com/install-windows/#RTX_4000_series_graphic_cards

set COMMANDLINE_ARGS=--xformers --opt-sdp-no-mem-attention --no-half-vae --opt-channelslast --disable-model-loading-ram-optimization