4 ms·

> Stable Diffusion? Well, that’s free-ish. Problem is, you’ll probably want to run it locally, which requires a really, really beefy graphics card. I was strugg

> Stable Diffusion? Well, that’s free-ish. Problem is, you’ll probably want to run it locally, which requires a really, really beefy graphics card. I was struggling to run it on a Vega56 - a GPU that goes for ~$150 used now - so I went out and got a RTX3090 for about $1,000. If you’re already a gamer with a GPU with 8Gb+ of VRAM you’re probably good, but for most people this is a bit absurd. Nit: 2 weeks ago there was an article about getting SD running on an iPhone. The app isn't very time- or battery-efficient, but it just barely works on my (previous-gen) device: https://news.ycombinator.com/item?id=33539192 https://news.ycombinator.com/item?id=33539192

ben_w 4y ago
I get results in 2 minutes on my iPhone SE. At least, when it doesn't stop execution part way through. When that happens, I have to change the seed, but it happens often enough that it sometimes needs a second change of seed to actually complete. Still, 2 minutes on my phone. I remember Bryce taking half an hour to render lower resolution images back in the late 90s.
orbital-decay 4y ago
Running on an iPhone-grade hardware is what they refer to as "struggling". You need a lot of iterations in any workflow, be it style transfer or text-to-image, so you want as fast generation and as much CUDA cores as possible, and up to 24GB VRAM for higher resolutions. More importantly, most of the power of this stuff is not just in inference, but in finetuning the model on your custom data, which requires 12GB VRAM for SD 1.4 at the very least, and ultimately multiple server-grade GPUs. And even more importantly, you need full control over your own box, not an iPhone app - as it's still a very experimental field which is only going to be more complex and hard to master. That said, you can just rent a GPU instead of buying it right away. It's not very expensive, and can be cheaper overall depending on your usage. It's also the only sane way to experiment with training as you need a lot of attempts. This stuff can really devour any hardware you can possibly throw at it.
- version_five 4y ago
  > you can just rent a GPU instead of buying it right away I agree, this is an underrated option and I'm not sure why it doesn't get more attention. I've been using stable diffusion on a cloud machine, and, say $20 will get you very far - probably thousands of images generated and a few sessions of fine tuning. It's way more affordable that its apparent popularity would suggest

ben_w 4y ago

I get results in 2 minutes on my iPhone SE. At least, when it doesn't stop execution part way through. When that happens, I have to change the seed, but it happens often enough that it sometimes needs a second change of seed to actually complete. Still, 2 minutes on my phone. I remember Bryce taking half an hour to render lower resolution images back in the late 90s.

orbital-decay 4y ago

Running on an iPhone-grade hardware is what they refer to as "struggling". You need a lot of iterations in any workflow, be it style transfer or text-to-image, so you want as fast generation and as much CUDA cores as possible, and up to 24GB VRAM for higher resolutions. More importantly, most of the power of this stuff is not just in inference, but in finetuning the model on your custom data, which requires 12GB VRAM for SD 1.4 at the very least, and ultimately multiple server-grade GPUs. And even more importantly, you need full control over your own box, not an iPhone app - as it's still a very experimental field which is only going to be more complex and hard to master. That said, you can just rent a GPU instead of buying it right away. It's not very expensive, and can be cheaper overall depending on your usage. It's also the only sane way to experiment with training as you need a lot of attempts. This stuff can really devour any hardware you can possibly throw at it.

version_five 4y ago
> you can just rent a GPU instead of buying it right away I agree, this is an underrated option and I'm not sure why it doesn't get more attention. I've been using stable diffusion on a cloud machine, and, say $20 will get you very far - probably thousands of images generated and a few sessions of fine tuning. It's way more affordable that its apparent popularity would suggest