V0.2 Important Notes:
PSA: Can produce mosaic censoring if "censored" is not in negative prompt, especially on penises. This is because the artist uses HARD mosaic censoring that I cannot edit out of the dataset.
Also, has a bias for lips and blush because the artist rarely draws characters without..
Trained with Danboru tags and Natural language, real pain in the ass to manually write/edit but, better prompt adherence.
Bumped to 4000 Steps.
Some recommended tags are : mature female, curvy, steaming body, sweat, plump, dark nipples/anus/labia
LoRA for the mommy enthusiasts
A style LoRA for Anima (by extension its finetunes) by CircleStone Labs.
This LoRA was trained on a curated dataset of 74 Illustrations from Toshi. All images were treated to remove text, any censoring and most symbols, then batch tagged with ConvNextV2 Huge Full and manually tagged for fixing/adding tags.
This Lora is best used at 1.0 strength, imo. Trigger word is very optional, but hey, its T0SHI.
Also this was trained on my chef's kiss AMD 7900XT, (under Linux with Ostris's Ai-Toolkit❤️)
Description
Main Settings: 2000Steps, AdamW8bit, 4Batches, [768,1024], LR 3e-5, WeightDecay 1e-2, Rank32, Repeats 2
FAQ
Comments (4)
Finally found someone that use amd gpu in this site
Yea, its pretty scarce ngl.
And less than two month ago I was still on a RX6800
Dabbling with local AI random stuff since mid 2024, never with an Nvidia card (colab doesnt count....)
It ain't slow If you never had something faster 🧠
Team red strong 🔴...ish
Uh what's the difference between v1 and v2 exactly?
Better prompt adherance because trained with natural language along danbooru captions, doubled step count with reduced learning rate.
It is way more flexible (usable), less hallucinating, and better cooked.
Thats important for exploiting the better text encoder of anima. For exemple the Gen pic was not great with v0.1. (Its also better for regional-ish adherance)