Maximum speed settings: (Requesting Testers!)

[ClashSAN]

Dear 3090/4090 users:

According to @C43H66N12O12S2 here, 1 month ago he is getting 28 it/s on a 4090.
I think he is busy but I would really like to bring attention to the speed optimizations which he's discussed in a long issue page.

If you have a 4090, please try to replicate, the commit hash is probably 66d038f
I'm not sure if he is getting big gains from increasing the batch size.

Anyways, if we use a sd finetuned model of 256x256, with a DPM++ 2M sampler at 6-8 steps, how many images would we be able to make?
See if --medvram in addition to --xformers will increase the speed, since it should raise your maximum batch size.

There are also some extra windows tips found at the bottom here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations Just don't try this on a very bloated system, or your results may not be good.

[leakedglass]

@ClashSAN Are you asking for only that model?

Trying the optimizations mentioned in that long issues page I have gained huge speeds. With my 4090, what was previously taking me minutes is now generating in seconds.

I did a new build around a 4090,... i9 13000, 64gb ddr5. I was only getting 1 - 2 it/s mostly around 1.05. Now I'm hovering around 17.90.

For example: (Run on a custom model trained from SD 2.1 model using Dreambooth and also using a embedding) to generate a 50 step DDIM at 960x960 with Batch size 8 was taking 6-7 minutes. Now it's doing it in just under a minute.

[ClashSAN]

17it/s according to some, is doable with 3090 as well. He was testing a regular 512x512 model as no 2.0 back then.
Your speeds on that commit version may be higher.