Hacker News Reader

Hacker News

Comment by nico | Hacker News Reader

Full thread

nico

Really impressive. Is that video sped up at all? Crazy fast!

View on HN

Not sped up.

Here's a live demo, but you need to register an account.

https://clipdrop.co/stable-diffusion-turbo

You need to provide an email address and click on a link. Emails from grr.la work.

Yes, it is "register an account", but it is the lowest friction I know.

Nope and this is just the beginning imagine a year from now or once the people that bring us LCM and controlnet and ipadapter start looking at the possibilities, not to mention fine tunes on turbo.

It's not sped up. I tried it locally last night and it was just as fast. Running on Windows 11 w/ a RTX 3090.

I just tried it locally with a 3070 and it was about 3 seconds per render. I'm far from great at this stuff and it was my first use of ComfyUI, so I don't know if that number could be improved on my setup.

On my machine with AMD RX 7900XT, it takes ~0.17s per image. Are you using SD Turbo Scheduler node?

Have you used the 7900XT with LLMs at all?

Yup, works just fine. Overall pytorch on ROCm 5.6 has been working very well. I'm impressed with how stable it is, given how much hate AMD driver stack has been getting.

same on my end, the biggest issue sometimes is finding the correct torch/torchvision version that go with different projects as AMD users are still rare. So, the problem is still the AMD ecosystem is pretty niche and will be pretty stable unless you pick a dev or wrong version, then it may happen to cause kernel panics! In any case, things looks to be improving, but rocm has a lot to support (xformers and other core ML libraries)

At the time I hadn't, but now I've used the sample image (the one with the fox in the bottle in the snow) to load up the workflow with sdturboscheduler, and it's still about the same time.

Edit: Even thought the UI says sdxl turbo, I notice that the command prompt is saying sdxl.

  got prompt
  Requested to load SDXLClipModel
  Loading 1 new model
  Requested to load SDXL
  Loading 1 new model
  100%|| 1/1 [00:00<00:00, 11.30it/s]
  Requested to load AutoencoderKL
  Loading 1 new model
  Prompt executed in 4.04 seconds
  gc collect

Edit 2: More info... I noticed that if I just change the steps, it takes less than a second to generate an image. If I just change the prompt, it takes 3+ seconds. I don't know enough about this to know what that means.

Doesn't seem that crazy 3070s can do 20 step 768px in like 6 to 9s

Reply on news.ycombinator.com