Just experimenting with checkpoint merging
I use articles: Crody's Guide and Model comparison
B3
Clip and poses: Unholy Desire Mix
VAE, colors, outlines and anatomy: Nova Cartoon XL
Face: Arthemy Western Art
and a little from other models: Toonify Zi-Illustrious MaturaComix
B2
Clip and poses: Unholy Desire Mix
VAE, colors, outlines and anatomy: Nova Cartoon XL
Face: PerfectDeliberate
and a little from other models: Toonify Zi-Illustrious MaturaComix
Description
Merge of:
Unholy Desire Mix + Nova Cartoon XL + PerfectDeliberate
FAQ
Comments (17)
v3.0 and v2.0 are 100% identical as proven be the identical hash code. You probably uploaded the wrong file?
I uploaded the latest version file to v2 several times, but for some reason the generator still using the old file, so I update to v3. That's why the hash matches. Thank you for noticing.
@vovkka I see. But now every image that is generated with v3 and uploaded to CivitAI will by automatically attributed to v2 instead, since their hash codes match and v2 has been on the page before...
@Lizardon1025 omg. I'll roll back the V2 file and hope everything goes well. By the way, thanks for you work on model comparison, it's really very helpful
Nice work! But please indicate the modells you used for the merge, if you not want trouble. The licence is a holy script what every creator should adhere to. Thank you!
The models used are presented in "About this version" and "Suggested Resources." But you're right, I should make the links more visible and be more careful with the licenses.
@vovkka Sorry man. My bad. I don't seen it.
Tbf, however, AI model licenses are an arbitrary set of rules we all imposed on ourselves, when the original datasets were all open-source. Open means open, after all, which means free and public for everyone to use, not subject to the limitations of whoever managed to merge a recipe that literally any user on the site could come up with.
In terms of base, non-merged models, the licensing authority comes down to whosoever owns the rights to the dataset (images, video, text, etc.), NOT the author who compiled the data. That's akin to saying the photographer who took all the school photos and brought them back to the shop has less authority over them than the guy who put them all in the yearbook. The issue, however, is it's nearly impossible to find, let alone credit, everyone with actual ownership over the dataset; as a result, the community has basically said "Well, SOMEbody should have ownership over it." This is not correct. Claiming a model as your own if you don't own the rights to 100% of the dataset is plagiarism. It's intellectual property theft, and, as such, the creator's licensing terms and requirements are unfounded and, more importantly, ultimately unenforcable.
@Chronos_Shinomori Sorry again. I don't want ot hurt anybody I respected someone elses work a much. Maybe I misunderstood the license.
@LumenFall No need to apologize. You're doing the right thing, after all. It's not an issue with you personally at all here, simply the ruleset with which we as a community have saddled ourselves. You're doing great work. It's always best to air on the side of caution when it comes to this sort of thing, so I would say that your policy is best, at least until more concrete rules and regulations are established.
[Meant to add this as a comment, but it posted it as a review. Positive either way.]
Best "experiment merger" ever. I took some prior metadata from past generations I had done to see how they would look/change with this checkpoint. They're wallpaper worthy. Prompts so far, I'd say are 90% on point. I love it!
I like this merger alot! Good job!
I'm sad to see that 1.0 got bumped off again. For some reason, I absolutely cannot get the same kind of quality in 2.0 or 3.0 that I had with 1.0. Especially when it comes to the face.
I have downloaded 1.0 and tried running it through Stable Diffusion, but the results are garbage and I can't figure out why.
@aliscarlet77158 Sorry. I thought no one was using version #1 anymore since no one had posted any pictures for a week... Now we have to wait until the next auction. Can you tell me which sampler and prompt you use to get the desired face quality?
@vovkka Yes I was using DPM ++ 2M Karras, and my prompt actually didn't refer much to facial details except for "red lipstick, full lips". The 1.0 generator just put out excellent quality for me regardless. I would be fine if I could replicate 1.0 locally but I must be missing inputs that you use here on CivitAI.