a1111-sd-webui-tome VS a1111-stable-diffusion-webui-vram-estimator

here how too install https://github.com/SLAPaper/a1111-sd-webui-tome

Good ole --opt-split-attention might give you some speed increases. Never treated me wrong. | Check out the optimizations page as well. --opt-sdp-attention seems to be faster than --opt-sdp-no-mem-attention, but less deterministic. Depends on what you're looking for. | You should also give xformers a whack. Here's a link to a precompiled wheel for python 3.10.9 and cuda 11.8. Some people have noted speed increases, some not. I've found Torch2.0 works just about as well by itself as xformers did on Torch1.13. Gave up on xformers for the night, but I'll definitely give it another test later. | Also, go grab the token merger extension. This thing cut my generation time by around a third. It's like witchcraft, I swear. No loss in quality/cohesion either.

VRAM Estimator Extension is a great little helper.

Either way, if you're using auto 1111, there's an extension that let's you benchmark your image generation capabilites and estimate what you can handle with available VRAM.

Look for the 'VRAM Estimator' extension. GitHub link

Batch size is directly limited by how much VRAM you have. The size of the batch dictates how many images to process at once. The more images being processed, the more VRAM you'll need. My 1060 6GB for instance can only run a batch size of 4 at 512x768, but I can run a batch size of 2 at 768x768. And a batch size of 1 at 1024x1024. The more images you want to generate at once, the smaller they'll need to be. | There's this extension to estimate VRAM. It runs a series of tests to find out the limits of your card and when it runs out of memory. It was a bit jank when I ran it, but it might be what you're looking for. | There's always the old tried and true method of "keep pushing it till it breaks". That's how I found my card's limits.