A style LoRA based on the art of kame_3t (@kame_3t on X), go check out their work. All credit for the style goes to them.
Soft pastel colors, gentle flat shading and thin lineart, mostly furry characters.
There are versions for Krea 2, MOSS, Anima and Anima 2.9B, so pick the one for your base model and read its notes (each version also has its own).
Krea 2
Start the prompt with @kame_3t, kemono, soft colors, thin lineart, clean lines,soft shading, and then describe the scene in full sentences.
Strength 1.0 (0.8 if you want it subtler).
Trained locally on Krea 2 Raw, epoch 9 of 12.
MOSS
Finally, one that came out right on the first try! The style showed up from the very first epoch, and the final epoch is the one I kept.
Start the prompt with @kame_3t, a kemono-style digital illustration with soft pastel colors, thin clean lineart and soft shading. and then describe the scene in full sentences.
Strength 1.0 (up to 1.2 if your checkpoint pulls too hard toward its own look).
Trained locally on MOSS Image Base, final epoch (12). No turbo or detailer needed: it looks clean on plain MOSS SFT at 35 steps and CFG 4.
Anima
The fast one: under 1 h 30 min of training. Trained locally on Anima Base v1.0 with Danbooru tags, epoch 9 of 12.
Start the prompt with masterpiece, best quality, @kame_3t, and then use regular Danbooru tags. Add soft colors, thin lineart, soft shading if you want the pastel look stronger.
Strength 1.0.
Use a finetune (tested on waiANIMA at 25 steps, CFG 4). On plain Anima Base everything comes out washed out, with or without the LoRA.
It sometimes draws a white border around the image (a quarter of the original art has one). Put border, white border, outside border in the negative if you don't want it.
Anima 2.9B: there's a dedicated v1 (Anima 2.9B) version, the same epoch 9 remapped to the 40-block layout of the 2.9B models. Strength 0.8 to 1.2. If your loader already remaps older Anima LoRAs (LoRA Manager does), the regular version works too, and that's also the way to go for 3.8B models.
Description
MOSS version, trained locally with sd-scripts on MOSS Image Base (int8) on a 16 GB RTX 5070 Ti, same dataset and captions as the Krea version.
Training: 144 images × 2 repeats × 12 epochs, lr 2e-4 with gradient accumulation 4 and a cosine schedule. It took about 2 h 40 min.
The style shows up very early on MOSS (already from epoch 1), and epochs 9 to 12 were all good. I compared epochs 6 to 12 against a no-LoRA control on MOSS Image SFT and MOSS-FET. Epoch 11 and the final one were almost identical (the cosine schedule barely changes anything at the end), so I kept the final epoch.
How to use
Start the prompt with @kame_3t, a kemono-style digital illustration with soft pastel colors, thin clean lineart and soft shading. and then describe the scene in full sentences.
Strength 1.0 (up to 1.2 if your checkpoint pulls too hard toward its own look).
No turbo or detailer needed: tested on MOSS Image SFT at 35 steps, CFG 4, euler / simple. Also tested on MOSS-FET and OneObsession MOSS SFT.
Things to know
Checkpoints with a strong look of their own (glossy, high-contrast anime) cover part of the style. Plain MOSS SFT keeps the soft colors best.
Like the Krea version, it can undress characters a bit. If you want them clothed, say so clearly in the prompt ("fully clothed", "kimono closed up to the neck").









