Locally with automatic1111, I’d say 10GB VRAM as a starting point to have a good experience (can do up to 1024x768 in a single go or a little less when ControlNet is involved). Though one can never have enough VRAM, when buying a new card I’d aim for 16GB VRAM to have some spare room for future models. Image generation takes about ~30sec. Upscaling is possible, but can take quite a while and results can be a bit hit and miss.
The plain StableDiffusion model is largely useless these days, go over to civitai.com and download something custom trained, they give far better results. ControlNet is another absolute must have and gives a lot of control over the resulting image (pose, 3d shape, sketch), along with a far superior inpainting, compared to img2img.
I have been playing around with SD for about half a year and it still blows my mind what kind of results you can get with rather minuscule effort. Though worth mentioning that the results are largely dictated by the AI, trying to get very specific results can be an absolute nightmare.
Runs fine with 1660, but that is about 20 sec for 512x512 and upscaling takes upwards of a minute.
If you want to run it online I suggest paperspace paid tier, not too big of a hassle to set up but you might have to wait a couple mins spamming refresh to get a better GPU, the instance can run for 6 hours, then it will be autoshutdown. Generally 2-4sec for 512 and 10-20 for 1024. Also, you will have to either download models every time, settle for only two or three models at a time or fork up a couple extra bucks for the permanent storage as base paid is only 15GB.
I even got it to run without GPU on a pure old i5 CPU with 8GB system RAM (not VRAM) paired with 32GB swap. SD1.5 takes 4-10 minutes per image, SDXL about 2 hours. But it works. With GPU its between 7 and 90 seconds per image, depending on model and settings.
I know on local llama I have been limited to 13b models
You can run llama.cpp on the CPU with reasonable speeds making full use of normal RAM to run much larger models.
As for 10GB in SD, I run into lack of VRAM quite constantly when overdoing it, e.g. 1024x768 with multiple ControlNets and some other stuff is pretty much guaranteed to overflow it. I have to reduce the resolution when making use of ControlNet. Dreambooth training didn’t even work at all for me due to lack of VRAM (might be possible to work around, but at least the defaults weren’t usable).
10GB is still very much usable with SD, but one has to be aware of the limitations. The new SDXL will also increase the VRAM requirements a good bit.
Add comment