CLIP SKIP = 1. Model was trained using CLIP SKIP 1.
I suggest using this VAE.
This model was trained and never merged. I suggest you merge it yourself. I did not train it as much as I wanted; but 146 hours of training using Free Colab is rough. For comparison sake, a person I know trained for 2x more steps than me in 12 hours.
It struggles a bit with img2img; you can and you should merge it.
Art style is a bit random but you can prompt for a lot of styles since is trained from a base model.
The main objective of this model is to train things popular anime models struggle with and experiment with highly curated captions/tags. Took me 100 hours to fix and tag the 5610 images used to train; I had to stop and code multiple features into my tools to make my life easier.
Other than that, just use normal Booru-style prompting.
If you want to use a LoRA, it's a good idea to use the LoRA at low weights.
Some of the tags I trained:
Adult characters with small breasts and flat chest; tagged breasts size correctly (in my own opinion, of course)
Also trained fishnets like fishnet tighhighs, fishnet armwear, fishnet top, fishnet pantyhose; used a very conservative type of textures since SD can't really deal with tiny squares that well cus VAE destroy it
Pearky breasts
Futanari, penis, testicles, bulge, erection under clothes, condom on penis, erection, flaccid, penis sizes, ball bra, cock ring, tucked penis, foreskin, phimosis, chastity, (these last 3 don't work well yet; needs more training)
Woman, pussy, cameltoe
Crotchless panties and cupless bra
Fantasy skin colors like blue, red, etc
Monster characters
Trained piercing and tattoos with correct tagging but doesn't seem to be working yet
Butt plug, heart butt plug, anal tail
Tentacle sex, consensual tentacles
Irrumatio, fellatio, cunnilingus
Also trained a lot of cum related tags, like projectile cum, and others
Pubic hair peek, female pubic hair, heart-shaped pubic hair (last one works but its kinda meh)
Clothing aside like panties aside, leotard aside and others; but it seems to be confused with crotchless panties most of the time
Doublouse stuff, kinda hard to pull off still
I'm sure I'm forgetting a lot of things but thats most of it.
Model is heavily biased towards MY own taste in artstyles and I avoided training young characters in.
Description
Struggles a bit with img2img and talking with some people I know, it seems like it needs more training to fix that.
FAQ
Comments (15)
I really like the asthetics in this model.. but the sample images really different in art style..
Is there a prompting guide how to get really simple sketches vs really detailed stuff? Or is it just random / biased into tags?
Can it do males as well?
Kinda random, the model is extremely undertrained for what I wanted from it. But I can't be bothered to train on colab anymore so I released it the way it is currently.
I did not test males at all, it should be able tho, since half the dataset is futanari stuff.
Also I was prompting with style tags in most images
So is this sorta meant to be combined with LORA's to drag it towards the style you want?
@DreamExplorer you should probably merge it.
Consider the model as undertrained... I don't have a GPU and I can't afford to be paying for cloud services.
Your best bet with it is to merge with something you like
The model seem great but feel like having the same issues than SDXL models, where we get random style with the same prompt depending the seed. A bit more stabilization could make this one a great model tho !
@legnaydal799 Yeah, but I don't want to spend another entire month training for 146 hours using Free Colab...
Unfortunately if you are interested in using this, you kinda have to merge with something else.
Just keep in mind this used to be the Anime model that shall not be named.
@Part_LoRAs I still really appreciate you sharing even a WIP model.. especially one you trained yourself. Thank you for the responses.
@Part_LoRAs Whatever thank you very much for sharing!
@DreamExplorer I had plans for training many other models, including a World of Warcraft one, but I don't think I have the willpower to boot up free Colab for another 146 hours to train...
I'm just making models while having my hands and feet tied.
@Part_LoRAs I find LORA's far more manageable.. I can't train models either. Too stressful.
If you ever find a super amazing mix formula with this model and some other stuff to push it more towards a consistent style.. you should upload it here as a new version. I'd be curious what you come up with.
@DreamExplorer I'm training it further but the best way to make it consistent is to merge it with something else.
That way it becomes more stable...
I could also merge a style LoRA in but I don't want to do that.
Nice!! Which model did you train on?
This started as the famous Anime model that shall not be named.
Trained that checkpoint for 146 hours on custom dataset as a learning experience.
@Part_LoRAs thanks! Will definitely try this one