A style LoRA based on the art of Movail (@Movail2), go check out their work. All credit for the style goes to them.
There are versions for Anima, Anima 2.9B, MOSS and Krea 2, so pick the one for your base model and read its section below (each version also has its own notes).
Anima (the best one)
Trigger: @movail at the start of the prompt.
Strength 1.0.
Regular Danbooru tags work fine.
Trained locally on Anima Base v1.0 with a cleaned dataset (signatures removed, so it shouldn't paint the artist's signature).
Anima 2.9B: there's now a dedicated v1 (Anima 2.9B) version, the same LoRA remapped to the 40-block layout of the 2.9B models. Use it with 2.9B models at 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.
MOSS
Trained locally on MOSS Image SFT, epoch 8.
Write your prompts in full sentences with @movail at the start.
Strength 0.9 to 1.0.
Looks cleanest with the MOSS Image Turbo LoRA (8 steps). Without turbo, use strength 0.9 and CFG around 3.
The RDBT detailer pairs nicely with it.
It's good, but honestly MOSS is still less faithful to the style than Anima.
Krea 2
Trained locally on Krea 2 Raw, epoch 16 (I continued from epoch 10 because the first run came out too weak).
Start the prompt with @movail, anime style digital illustration, cel shading, and describe the scene in full sentences.
Strength 1.0 (up to 1.2 if your checkpoint pulls toward realism).
Tested on Neutrino v2 at 8 steps and CFG 1.
It likes to add sharp teeth, like the original art. Write "small, rounded teeth" in the prompt if you don't want them.
Krea leans toward realism, so the style words at the start matter. Anima is still the most faithful version.


Description
Krea 2 version, trained locally with musubi-tuner on Krea 2 Raw (int8) on a 16 GB RTX 5070 Ti. This one fought back.
The training story
The first run was 10 epochs overnight (~4 hours). The style was there, but so weak that I needed strength 1.4 to see it.
So I continued from epoch 10 for 6 more epochs instead of starting over. On the way the GPU driver crashed once and the training slowed to a crawl until I freed some VRAM, but in the end it ran in under 2 hours.
In total: 16 epochs, 3200 steps, around 6 hours of GPU time plus a couple of failed attempts.
I picked epoch 16 after comparing epochs 9 to 16 against a no-LoRA control, on several characters. It was the most stable and the closest to the original style.
How to use
Start the prompt with @movail, anime style digital illustration, cel shading, and then describe the scene in full sentences.
Strength 1.0 (up to 1.2 if your checkpoint pulls too hard toward realism).
Tested on Neutrino v2 (Krea 2 Turbo based) at 8 steps and CFG 1.
Things to know
Krea leans toward realism or western cartoon. Without the style words at the start it can come out semi-realistic.
It likes to add sharp teeth, like the original art. If you don't want them, write "small, rounded teeth" in the prompt (at CFG 1 the negative prompt does nothing).
Sometimes it adds a white sticker-like outline around the character or a frame around the image.
Honestly, the Anima version is still the most faithful to the style, so if you can, use that one.









