Speed and optimisation

[Njaecha]

I'm starting this as Q&A because its mainly a question I've been wondering about:
Why is there such big speed differences when generating between ComfyUI, Automatic1111 and other solutions? And why is it so different for each GPU?

A friend of mine for example is doing this on a GTX 960 (what a madman) and he's experiencing up to 3 times the speed when doing inference in ComfyUI over Automatic's.
I, on the other hand, are on a RTX 3090 TI and inference for me is 4 to 6 times slower than in Automatic's. I'm using xformers on both systems and were testing with the same settings.

And then in addition, complicated workflows with more advanced conditionings get even slower (which is to be expected).
I'm wondering if other people experience similar things and if anyone has ideas on why it behaves like that. And how we can make it better (on strong cards).

[comfyanonymous]

Try using an fp16 model config in the CheckpointLoader node. That should speed things up a bit on newer cards.

It should be at least as fast as the a1111 ui if you do that. If it isn't let me know because it's something I need to fix.

[PospoSK]

Found this fix for Automatic1111 and it works for ComfyUI as well.
https://www.reddit.com/r/StableDiffusion/comments/y71q5k/4090_cudnn_performancespeed_fix_automatic1111/

I have RTX 4070Ti a had this gains with update to cudnn8800 (Model ChilloutMix, 512x512, 20 steps, CFG 8, Euler):
Batch 1 ... 8.5it/s to 17.6it/s
Batch 4 ... 2.6it/s to 4.5it/s (10.4it/s to 18it/s)
Performance wise are both equal now. And I am happy :)

[comfyanonymous]

I think this can be closed because you can just use the CheckpointLoaderSimple now to get optimal performance.

[cobrahnaf]

It took me like 1.5 hours to generate this image, and yeah the image is not what I expected it to be, it's so blurry.

[Femmyscout]

cfg 100.....