Cutting Edge
This resource is about experimental checkpoints for cutting edge model development (SD1.5 & SDXL1.0). It is currently focused on anime style.
Description
FAQ
Comments (19)
Interesting style, I like.
The style is kind a good anime illustration, but the checkpoints seem to have some issues with prompt adherence. Need to address this.
@Hevok It might not be THAT specifically, it depends largely on clip and your order of prompt, i'll take a gander next time i'm on AI, now that the challenges are done i'm trying to take a bit of a slower break trying not to go back on for a few more days XD
which vaes would you recommend for both?
will test both checkpoints on my prompts ;D
I use the built-in VAEs.
@Hevok there's something that i noticed pretty quickly with both of your checkpoints:
both totally hate weights. every time i use a weight higher than 1, i get images without any style suddenly or monochrome images or freakshows. it's super annoying to be honest.
i have prompts that need with weights in positive and on your checkpoint, sadly, they are worthless.
@stapfschuh Oh, interesting. I created these checkpoints in a totally different experimental way with a newly crafted code base. I did actually not test them with any weights on tokens perhaps. Just text descriptions and tags mostly. Let me investigate and try to fix it if possible. Thank you very much for the feedback!
@stapfschuh I checked back the tests I did and I realize I was able to use weights in the prompts just fine for both SDXL1.0 and SD1.5 versions in ComfyUI. You used Automatic1111 WebUI right? I would like to replicate the issue so I can troubleshot it before proceding with further development.
@stapfschuh When i attempt to download the version from Civitai the file name is suddenly so weird. "edgeSD15SDXL10_xl10V1.safetensors", but it should be "edge_xl_v1.safetensors"
@stapfschuh Could you please provide an image with generation data of the problem? I cannot replicate it so far on either ComfyUI nor Automatic1111.
@Hevok you are right. i use a1 (short for automatic1111 webui).
when i downloaded it, i realized that strange name at first and thougth "what a mess!" good to know that it wasn't on purpose and not even your fault :D
should i rename them? but i don't think the filename does change anything.....
@stapfschuh File name should not matter at all. Civitai just does some weird thing of trying to incorporate the model title and version name. I intend to raise this issue in the next office hours.
@Hevok uploaded 2 images here. normally those negative prompts force perfect hands and that 1 positive-prompt (normally i don't use them together. it was just for testing, but both work together. normally) generates good looking hands. your checkpoint keeps ignoring all of them.
@stapfschuh Oh, so it is very subtitle. Prompt adherence issue maybe. I was expecting some kind of extremely messed up images. The images just don't show any hands at all. Either by putting them in the pockets or out of frame.
@stapfschuh True, hands are not good. I tried your prompts:
SD1.5: https://civitai.com/images/4928779
SDXL1.0: https://civitai.com/images/4928858
Let me try to fix this. Thanks!
there is a second point that i found out about your sd (not xl) checkpoint. it seems it can't handle resolutions higher than 512x768. because, using this resolution, some of my prompts work again, but changing it to 512x1024 is ending up with your checkpoint basically saying "nope!", because nothing is working anymore.
So it follows the prompt better at 512x768? I generate at 768x1024 with this SD1.5 checkpoint. Keep in mind that Stable Diffusion 1.5 base model was trained at 512 resolution. Going beyond this may lead to artifacts. 768 is usually said to be supported though.
@Hevok there is something that i should give you for tests on your own:
put (defaults) in your 2 checkpoints and tell/show me if it generates upright standing characters with all 10 fingers intact. using other checkpoint, i get exactly this as a result, but using your checkpoint, there is not even a slight reaction. nothing. it's a nightmare for me when i try to "create/find" new prompts to see no reactions/results.....
IDK, works perfectly for me at 1024x1344 using Fooocus.