Hacker News Reader

Comment by smrtinsert | Hacker News Reader

I have a 6gb 1660ti, barely holding on. Is a new 12gb card good enough for now, or should I go even higher to be safe for a few years of sd innovation?

View on HN

I'm using it with a 2070 (4 year old card with 8gb vram) and it takes about 5 seconds for a 512x512 image. It's been plenty fast to have some fun, but I think I'd want faster if it was part of a professional work flow.

What settings? That seems faster than expected.

It was the defaults for the webui I used. Faster than I expected too, but the results were all legit.

Edit: Got home and was able to double check. It's actually a solid 10 seconds per image with the following settings: seed:466520488 width:512 height:512 steps:50 cfg_scale:7.5 sampler:k_lms. Still quick enough for some fun, but could be annoying if you're need to do multiple iterations a minute.

Two minutes with my 1060, sadly.

im on my 2020 macbook air m1 ... 512px image takes 2-3 minutes :(

The GeForce 4000 series is about to release and should make Stable Diffusion wayyyyy faster based on related H100 benchmarks posted today.

It sounds like there's forks that are able to work with <=8GB cards. And I'm not sure but I think the weights are using f32, so switching to half might make it yet easier still to get this to work w/less memory.

But yeah the next generation of models would probably capitalize on more memory somehow.

People have reported that this repo even works with 2gb cards if you run it with --lowvram and --opt-split-attention.

Yes, the amount of VRAM doesn't seem to be as much of a limitation anymore. However, processing power is still important.

How is M1/M2 support for SD? Is there a significant performance drop? Presumably you would be able to buy a 32GB M2 and be future proof because of the shared memory between CPU/GPU.

In my setup at least it runs essentially in CPU mode since there is no CUDA acceleration available and metal support is really messy right now. So while quite slow I don't run into memory issues at least. It runs much faster on my desktop GPU but that has more constraints (until I upgrade my personal 1080 to a 3090 one of these days).

There was a long thread last week. It’s honestly pretty good if you follow the instructions. 30-40 seconds/image.

Yeah, I followed the instructions on a M1 Macbook Pro (Monterey 12.5.1) and it worked without extra effort. 30-40 seconds per image. I have 32GB but image generation doesn’t even use half of it. The hard part has been to generate prompts that do what I want.

I recently switched from a CPU-only version to this repo release 1.13: https://github.com/lstein/stable-diffusion

The original txt2img and img2img scripts are a bit wonky and not all of the samplers work, but as long as you stick to dream.py and use a working sampler, I have had good luck with k_lms, then it works great and runs way faster than the cpu version.

Works great on 32gb ram but I'm honestly tempted to sell this one and get a 64gb model once the m2 pros come around. This is capable of eating up all the ram you can throw at it to do multiple pictures simultaneously.

Reply on news.ycombinator.com