@SG_161222 CivitAI says otherwise on DL-Page :)

@SG_161222 Yeah - this would really help.

thanks @SG_161222 will check it out today.

@SG_161222 tried but could run it for some reason...it seems to result in OOM issue with diffusers library on an A10G GPU. I can run base sdxl fine. Also, would be better to have safetensor version as that's the default way even for diffusers.

Let's see yours ;)

The one number you didn't include, VRAM, is the only number which matters for that really.

I was able to use it with 8 GB of VRAM by adding the --lowvram flag when launching A1111. ComfyUI figures this out by itself and didn't need any manual config for me. Any less than 8, I wouldn't bet on it working.

I finally made the switch to comfy after this exact issue happened for a1111 time and time again - all working correctly on comfy. I have a 16gb ram, 8gb vram on a rtx3070 graphics card just for reference (asus tufdash F15 laptop. Try out comfyui. It's not complicated to pick up at all, and it works so very much more consistently than a1111.

@Maeyanie I have 8gb of VRAM. I'll try the lowvram flag. Took me quite awhile to understand the basics of A1111, I don't want to learn another interface right now.

Edit: the lowvram option is not workable. It takes 45 minutes for one picture? Sheesh

I have seen this image on my own model page as well. I think people have some research to do, it should work fine. I believe people are using resolutions that are too high to handle for the VRAM, that is all. I am doing 1700px+ renders here with A1111 and no problem.

@JohnRohan I used to have 6gb vram and I used medvram which seemed to work. From what I understand it shouldnt have worked but it did for me. That was before any of the XL models tho so probably wont work but might be worth the try. Also if you drag one of your images into comfy it will copy the setup for that image, you can also download some other people's workflows for it and its pretty much plug and play if you do it like that

@olivetty I was only trying to create a single test image at 512 x 768. Again, I have had no memory errors on any other models even at higher resolutions.

@JohnRohan That's weird, I'm using a 3080@10GB with --medvram A1111 1.6.0 flags on my PC (R7 5800X + 16GB RAM + 2TB M.2 NVMe SSD) and doesn't have errors even training or generating with SDXL.

Also in my laptop (R7 6800H+16GB RAM) with a 3050@4GB, using --lowvram flag, I'm also having good results at generating txt2img and img2img SDXL images. Not training.

Try to use ComfyUI, at least it doesn't use RAM/vRAM instantly like A1111 so it could help.

@JohnRohan try adding --medvram-sdxl

@JohnRohan very strange indeed, hope ypu find what is wrong, in the meantime, try this; https://www.reddit.com/r/StableDiffusion/comments/168sad2/comment/jyy63za/?context=3&share_id=bsS6Xbe8VB9N5vGWe2OAQ

I have a 6GB VRAM 2060 laptop and it works great with this or any other SDXL model, including Lora and controlnet... If I use Comfy. Auto or similar is useless, it "works" but is 6-7 times slower. In comfy I can generate a 1024 by 1024 30 steps in less than a minute, with Auto it takes 7 minutes....

Try "Eular a" sampler for eyes.

I agree 100%. This SDXL model is the best for photorealistic generations so far and the one I use as my daily driver.

What do you mean by relearning your prompts? how would you do that? sample submitted images from this page?

StableDiffusionXLPipeline.from_single_file(path_to_model)

@SG_161222 I'm already on it it's beautiful. Best model really. I cannot wait for the next update.

you need an heavier negative prompt. That one suggested isn't enough, particularly in complex scenario. Search other pictures and try to copy some negative prompts that include something about mutation... search also in other models.

if you are using hires fix make sure your denoise is 0.4< or there abouts