Hello everyone.
I decided to finally rename and reupload this model, and clean the gallery as well. This is the same model as before, the only difference is I unbaked the VAE file. My reasoning is maybe someone prefers raw looks. Or you already have this VAE file. Or you wish to attach a different one. The default VAE is the file from schoolmax2.5d, renamed for convenience. Baked VAE version is available as well.
I removed the previous wall of text cause nobody would want to read it. Basically, this model tries to make 2.5d images with painted eyebrows and pubic hair. Sometimes it works, sometimes doesn't. I think it's a decent first step. It can't do hands, it can't do background people (seriously, don't try that), and it's better in portrait mode than in landscape (unless you're doing a landscape, of course), but what's new? It is equally good for SFW and NSFW. Thanks to users lynx and ritcher1, I found this model is extremely sensitive to weights in prompts. So maybe it's not as niche as I thought.
The only drawbacks I found here is that the resulting merge doesn't like positive textual inversions or LORAs. Anything I tried to add, even Master Corneo's POVs, cause it to lose 2.5d effect completely, reverting it into 2d model. What's weird as I never encountered this behavior even on fully photorealistic models. Negative inversions are fine, however.
Now, for settings.
In my tests, DPM++ 2M Karras was the best sampler. Steps from 20 to 32, scale about 10. As of recent, I started to heavily trim my prompts, eliminating useless buzzwords like masterpiece or unity 8k wallpaper from positive part. For negative, I tend to use negative embeddings like bad_prompt_version2, Deep Negative, EasyNegative, and Unspeakable Horrors. on top of basic (worst quality:1.4, low quality:1.4). I increased clip skip to 2 and I like the results. Face restore seems to be not very compatible with this merge, it curses faces. Photo realism tag works extremely well. Eyes are still cursed sometimes, especially when steps are too high. Fingers are almost always cursed, nothing new here. I have xformers enabled so your results may vary a bit, mostly in colors.
Thanks to DiaryOfSta (AniDosMix) and MJXXZ (schoolmax2.5d). Every rights and credits remain by their creators and/or owners. I take no credit beyond the basic mention. Both original authors disabled some kind of commercial use so I'll just disable it altogether. Speak with them, if you have to.
Description
First release, basic version. VAE file is loaded for pictures in gallery, and baked in the model file.
FAQ
Comments (25)
Yes, where is the model SchoolB ? Is it an internediate model, made during the building of the final mix ?
It's just the cringey working title for this merge, sorry for the confusion.
So that render based on just SchoolB, can be made with the final mix ? That render is totally different.
@ritcher1 Yes, I think this prompt reverted it to the base schoolmax2.5d model. Look there, you'll easily find the swimming girl in comments. I was surprised as well. But as I wrote in my introduction, I needed this for a very niche purpose. This merge does it, while the base one doesn't. If you're interested in the style of this picture, you're probably better off with schoolmax2.5d. It's insanely good for this.
I've added the same prompt but made with current name. Then I switched the dimensions. Then remade both pictures with something close to my OC, leaving everything else untouched. Maybe this'll show the range better. The old picture with swimming girl I'll leave until tomorrow so anyone can compare for themselves.
@Minamikaze
Ok, I've seen the two renders, thanks. In my opinion the render made with SchoolB is better than the other versions, more detailed, more realistic. I think that your intermediate SchoolB is a good blend on its own. Any thought about making it alive on CAI ?
EDIT: the following was written before reading your previous post. Thanks for the renders comparison.
Don't know... the render made with your schoolb is better than the similar one in schoolmax2.5d. Can you apply the same (your) prompt to the final mix ? If your blend is so flexible it would be even more good to have it.
@ritcher1 Yeah, refresh the page. My sleepy brain made me to forget to disable additional style for negatives so it got fed some others (and some details were changed as a result) but it's undeniably the same render. And I'm going to get some sleep... Tomorrow I'll add a couple of other pictures in different scenarios.
@Minamikaze Ok, that's it ! Well, your final blend is more flexible than what appears from the first batch. Good idea to add other renders to exploit its full potential. Thanks for your help.
@ritcher1 There were no intermediate model, as I wrote. It is the same file. The first picture was made yesterday, and the model got renamed today. That was the only change. Any minor differences between the old portrait and the new portrait are due to me slipping in more negatives by mistake. You may confirm this easily by copying the outputs into your machine (but do install negative inversions first, they had to do SOMETHING for the final result, I guess).
@ritcher1 No, thank you for your feedback. I would never see I used the picture with an obsolete name otherwise. Confusing people was never my intent so I'll fix that, first thing in the morning.
@Minamikaze Ok, we'll see tomorrow. Btw, are you using controlnet for your renders posing ?
@ritcher1 I've identified the reason behind these dtastic changes. It's weight in prompts. Uploaded a bunch of pictures differing only in one line being progressively weighted forward, and I think you'll notice the difference right away. And no, I'm not using anything beyond the basic webui.
@Minamikaze Thank you for your efforts. I'd now get rid of the testing images (just 2 is enough), since the gallery is a bit too crowded.
Just a couple more questions:
- which was the " one line being progressively weighted forward" ? The prompts are a bit filled.
- can you try to recreate the questioned image (the schoolb girl underwater), same prompt, with the pruned fp16 version ? How much the quality will go down ?
@ritcher1 Yeah, eventually I'll do just that. Starting with removing three horrible DDIM images. But not today.
@ritcher1
1. I did one of the girls. Probably the basic one, but can't remember. And I don't think there were any differences at all. Next week, I'll try to separate the pruned model, doubling everything for clarity.
2. The bit about "extremely long shiny silver hair". Nothing else was changed between brown-skinned swimming girls. Right now, I'm generating other pictures using this model, and I had to use plenty of weighted bits. It resulted in the face becoming as flat as it was in AniDosMix_B. So, basically, the more weights you use (proven for positives), the further you're getting from schoolmax2.5d. Most of the pictures in my gallery are somewhere in the middle, hence the difference between that swimming girl (it has almost nothing weighted in it) and other pictures.
@Minamikaze Perfect, thanks again for the clear explanation. Now I'm going to visit your gallery ...
Interesting model, but can you share a smaller 2gb pruned fp16 version ?
Hmmm. I'll see about that.
I@Minamikaze Ok, let me know whenever you do it. I'm getting low on free disk space after hundreds of downloaded models. And I'd like to have this one too.
@ritcher1 Done. I've tested it on some pictures, seems to be working fine.
@Minamikaze Thank you very much !
I'd rename the pruned fp16 version "anidosschoolb_10_pruned_fp16" just to reminf the users what they got.
@ritcher1 Maybe later. Uploader bot was very buggy for me today, I'll let it blow some steam off. Thanks for suggestion, though.
@Minamikaze Another suggestion: I'd separate the renders in the "Versions" section, between the full and the pruned version, so to make evident the difference, using the same prompts. In theory the results should be almost the same, but still ...
@ritcher1 Yeah, it actually is the main case I'm not ready to do it right now. Truth is, I already did it the first time, but the uploader caught a nasty bug, forcing me to redo everything. So I just put everything in one place as it's easier and called it a day. Next time, I will be better prepared to do it faster.
@Minamikaze Ok, no problem.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.