2y ago

My 3070 cannot seem to run SDXL

I have been running 1.4, 1.5, 2 without issue - but everytime I try to run SDXL 1.0 (via Invoke, or Auto1111) it will not load the checkpoint.

I have the official hugging face version of the checkpoint, refiner, lora offset and VAE. They are all named properly to match how they need to. They are all in the appropriate folders. When I pick the model to load, it tries for about 20 seconds, then pops a super long error in the python instance and defaults to the last model I loaded. Oddly, it loads the refiner without issue.

Is this a case of my 8gb vram just not being enough? I have tried with the no-half/full precision arguments.

17 comments

I had issues before I updated A1111. Do a git pull in the A111 directory and try again.
- I have mine auto set to git pull each load. Can you confirm, since you have it working...what files do I actually need - is it just the base, refiner, lora and vae?
  
  I’m not an expert, but what I read said that you use SDXL by first using txt2img to generate an image using the base checkpoint, and then you send that image to img2img and use exactly the same prompt there with the refiner checkpoint.
  That makes for a longer workflow than I’m used to, so sometimes I just use one or the other in txt2img and see what I get. Sometimes I forget to change the model when I switch between img2img and txt2img, too. I always seem to get results of similar quality when I use just one of the checkpoints.
  It should be interesting to see what people come up with training their own checkpoints off of SDXL, though.
3060 here, it might be the vram. SDXL eats a lot of it (and if you had say the vae in the wrong spot it would output very wrong images) so it might be that either 8gb aren't enough, or maybe they aren't enough with the resolution of your screen plus whatever you are running, like the browser.
Or, OR: the checkpoint is corrupted. I had that happen a couple of times in the past and the whole huge error with loading of another model was what happened.
Nvidia has problem wih newest drivers, auto 111 give out of memory, comfyui works smootly with your card.
- Is there a driver version that I can downgrade to? I just cannot do Comfy. I gave it a good try. I have used GRisk, Auto, and Invoke and liked them all - I just cannot get used to Comfy.
  
  https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/11063
  This is the thread, they explain better the issue.
I'm not sure why, but I have 8GB vram and my experience with this has been the same as others who describe that SDXL will not run with Auto1111 but it will work with ComfyUI. So I think this is not purely a vram issue.
- Auto1111 might be trying to load multiple models at the same time, which it does not have room for.
- Yeah it's very odd. I tried comfyUI. I but the interface just doesn't click with me.
  I keep waiting for invoke AI to have an auto installer for that model but they are still only offering the SDXL .9 and I don't have a token for that model.
SDXL is very memory hungry. Most base models are around 6-7 GB, which doesn't leave much room for anything else.
- Thanks. Oddly enough, the most recent release of InvokeAI fixed the problem I was having. My 8gb 3070 can run SDXL in about 30 seconds now. It seems to take a little bit to clear everything in-between generations though. I want to move up to a 12/24 gb GPU, but am waiting/hoping for the price crash.
  
  Awesome!

17 comments