This LoRA is based on the art style of kangwolf ( https://x.com/tony07734123 ), go check out their work. Thick white sticker-like outline, flat cel shading, sleepy half-closed eyes and cozy everyday outfits. The dataset is mostly furry characters, but it handles other species just fine.
Krea 2 (v1)
Trained on my own PC on the Krea 2 Raw base. Epoch 13, strength 1.0, and write your prompts in full sentences like MOSS.
About the trigger: the original art has the artist's "Kw" signature in the bottom-left corner. I cleaned the watermarks from the dataset, but some traces remained, and the trigger @kangwolf ended up tied to the signature. If you use it, the LoRA tends to draw the "Kw" mark in the corner. My advice: leave the trigger out. The style stays just as strong without it and the signature is gone (at most a very faint trace in the corner that a quick inpaint fixes).
Tip: to keep it from going realistic, end your prompt with something like: "The character has a thick white outline and flat cel shading, and the image has clean edges with no text."
MOSS Base (v2)
Retrained on MOSS Base instead of SFT, with a recipe I pieced together from the metadata of MOSS LoRAs that work well (rank 32 / alpha 32, lr 2e-4, gradient accumulation 4, cosine). Epoch 9, strength 0.9, prompts in full sentences. This one works on its own: no turbo, no detailer, just 25-30 steps and CFG 4. Turbo LoRAs are optional now if you want speed. Same as Krea: leave @kangwolf out of the prompt.
Recommended strength: 0.9
I'm still learning MOSS. It's a new model and there isn't much info yet on how to train it, so expect more updates as I figure it out. v1 stays up for anyone who likes it with turbo, and if you have training tips I'd love to hear them in the comments.
MOSS SFT (v1)
Trained on my own PC on MOSS SFT. Epoch 14, strength 1.0, prompts in full sentences. Honestly, I still haven't figured MOSS out: on its own it doesn't do this style justice, but with the MOSS Image Turbo LoRA ( https://civarchive.com/models/3006599/moss-image-turbo-loras?modelVersionId=3420002 ) and RDBT | MOSS ( https://civarchive.com/models/3007506/rdbt-or-moss?modelVersionId=3457029 ) at 1.0 it looks great. Same as Krea: leave @kangwolf out of the prompt.
Also recommended for MOSS: the turbo LoRA from MOSS/Kromoss General NSFW + Aesthetic/Style + FurryPlus ( https://civarchive.com/models/3031882/mosskromoss-general-nsfw-aestheticstyle-furryplus-normalturbo?modelVersionId=3444459 ), which pairs really well with this one for furry characters.
If you want the best result for this style, go for the Krea 2 version.
Anima (v1)
Trained on my own PC on Anima Base v1.0 with tag captions. Epoch 12, strength 1.0. Prompt with tags (mixing in natural sentences works great too) and add "white outline, outline" for the sticker border. Here the trigger works fine: put @kangwolf at the start and "watermark, signature" in the negative prompt.
Anima 2.9B (v1)
The same LoRA remapped for the 2.9B models (40 blocks instead of 28), no retraining needed. Only use it with 2.9B models, strength 0.8 to 1.2. If your loader already remaps Anima LoRAs (like LoRA Manager), the regular version works too.
Description
First Krea 2 version, trained on my own PC.
56 images (3 repeats), trained locally with musubi-tuner on Krea 2 Raw (int8), rank 32, lr 1e-4, 18 epochs. About 4.5 hours of GPU on an RTX 5070 Ti. I went with epoch 13.
Heads up about the trigger: with @kangwolf in the prompt the LoRA tends to draw the artist's "Kw" signature in the bottom-left corner (traces of it were left in the dataset). Without the trigger the style is just as strong and the signature is gone, so I recommend leaving it out.
How to use it
Write the prompt in natural sentences, strength 1.0. End the prompt with something like "The character has a thick white outline and flat cel shading, and the image has clean edges with no text." to keep it from going realistic. Tested on Neutrino v2.







