The project is closed. Thanks to the 10 people who supported the last two models, 6,500 buzz's. This was enough to train three LoRa models, which is, of course, impossible for further model training.
Description
FAQ
Comments (17)
Any plans for an fp8 version?
Why do need it? I have a 4070 laptop with 8GB of Vram, the rest is DDR5 RAM, and the speed at this resolution is 3 sec/it. When I run the FP8 version, the speed remains unchanged, but the number of errors in anatomy and following the prompts increases.
https://github.com/qskousen/ggufy
Is as easy as just drag & drop and supports fp8.
@Viennar because many of us don't have 8GB RAM even and still use ZIT, plus it takes a lot less disk space. it's an issue, I currently have 28 ZIT models to test, and just spent a week throwing out 500 gb of sdxl, flux and sd1 models. So it would be nice, I don't download these since I can't make room for them, but I'd love to try it. Plus the download time on CivitAi, it depends where you are, but I have gigabit Internet and still 6GB takes like 20-30 minutes to download.
@Kerstal Ooo thanks - that looks useful!
Very good model but lets be honest: like most ZIT models, it absolutely can't do vaginas.
Vaginas, yes, can be achieved in 2-3 attempts if the body position is non-standard)
solid model for sure, vaginas or any southern regions can be achieved with other methods :)
@MrGhost_27 Happy to get a few pointers. All the LoRAs I've tried change the faces enormously.
@Viennar Hmm... none of your examples show a correct one though
Yes use a lora most models are not there yet. But there are like 5 different loras for it among the first 20 for ZIT. It's one of those loras you really should use even if you hate loras.
Tertium
Is fantastic and does not prevent varied faces being generated via {desc|nDesc} style prompting, it's fantastic.
Thanks for making this. how about one to make me use most of my 32gig of VRAM?
Hello, there is no difference in quality between FP16 and FP32. Perhaps the problem is that the Loras training on Civitai makes them look like FP16, or I don't know the reason.
Is there additional trained CLIP for the checkpoint? I'm using ZIT CLIP from CivitAI.
I've tested Tertium and even with really simple prompts for example missionary sex I usually get something which should probably captioned as grinding or frottage. Usually checkpoint stops to the point where scene is "implied" but not actually happening. I'm using sex act wording from Danbooru tag groups but with natural language: https://danbooru.donmai.us/wiki_pages/tag_groups
e.g. simple test prompt "blonde male. woman with short black hair. male does missionary sex with woman." Positioning works somewhat with "laying on her back" and "spreads legs" and so on but it really do not much to improve situation.
So I am pondering if CLIP needs to be trained too for new wording and concepts.
On positive side, there are not (at least much) vaginas in place of penises so that part is better than Secundo.
Hello, the model hasn't yet been trained to use position tags like in Illustrious. Currently, for the messaging pose, the scene should be described in full: "Photo of a girl lying on a bed, her legs spread, a penis inserted into her pussy, vaginal sex. POV, first-person view, the girl's legs are held by a man, the man's stomach and legs are visible." Instead of "lying on his back," it should be written "lying with his stomach up," otherwise the back will be shown. There's still a lot of work to be done.
Whoever wants to really "fix" ZiT must address the issue with those high pussy slits that completely ruin realism.
ZIT could still learn normally. I've already spent over 50,000 buzzes on experiments, and it's unclear how much more will be needed. I hope I'll achieve satisfactory results.