RDBT [Anima]
Mid scale finetuned + guidance distilled.
I use it as a starting point to stack more style LoRAs.
See this page for update log. Random experiment, random quality. New version != better version. Feel free to leave feedback.
See this page for original LoRA (update more frequently, probably).
Sharing merges using this model is not allowed. Known model thieves: NukeA.I (closed-weight merged model on tensorart),
This model is based on
ym: AnimaYume (hf link) (civitai link). Has latest dataset.
b,p: Anima pretrained (hf link)
Usage:
Settings:
CFG scale: 1~4. This model has been guidance distilled. You can disable CFG (CFG 1) and run the model 2x faster. Cover images are without CFG for demonstration.
Steps: 24. Guidance distillation != step distillation. If you need low steps (8~12). Try to add 0.2~0.5x turbo lora.
Prompt
Always specify style, or use a style LoRA. Otherwise, you will get random/mixed style. This model does not provide overfitted default style. This is a feature, not a bug.
Quality tags:
It's recommended to omit all the quality tags, or just keep the "masterpiece", if you're not confident. Omitting those redundant tokens allows LLM to pay more attention on other words.
Quality tags have been reinforced during distillation. Thus they don't have noticeable effects. Same as negative tags. If you use cfg, there is no need to dump "score_1, blurry, worst quality, jpeg artifacts, extra arms,... x100 words" in your negative prompt. Those things have been distilled out.
Training settings:
~10k images finetuning -> guidance distillation
All captions are NL from Google Gemini.
Optimizer: adamw, constant lr 0.00002.
LoRA rank/alpha 24.
Guidance distillation target CFG 4.
Block 0-2 and adaln linear layers are skipped.
Description
FAQ
Comments (11)
Without picking a specific style, the generated characters lean into a semi-realistic AI look (tested 10+ times). It works perfectly for anime styles once I add the right style tags. Overall it’s pretty good, gonna go test out more style tags next.
Yes, there is no default style. Without style conditions distilled model will give you the "ultimate AI style" that averaged every style. Which looks really bad.
Bossman producing banger after banger
0.29 produces some great results when you're asking for something generalized and well understood but I feel like prompt adherence went out the window for fringe content. Anima as a whole is really good at mashing concepts together using natural language, but throw in some simple contradictions ("A solo illustration of a girl with a penis") and it struggles to understand.
Try and make sure that you specify the style or artist at the beginning of the prompt to lock in that style. Otherwise it can be a little bit wild with all Checkpoints especially the base one. pretty much all of them unless it's specifically stylised towards a specific style.
I know you are not really supposed to use the turbo lora with this checkpoint since it's already turbo/distilled to begin with... buuuut I did anyway and the results are honestly very interesting. A lot less detail but VERY strong 2D anime styles are possible with very stable composition. I kind of love it. Posted some pics.
I've tried it with the Cosmos Predict 2.5 DMD2 LoRA too and it works well.
Like the other person, I tried v0.29 RDBT Checkpoint, Cosmos Predict, and Anima Turbo LoRA all together with 4 steps. Pretty interesting stuff!
V0.29 exceeded all limits in flash-banging with 3D looks/shading
Further we are from V0.24 the worse 3D it gets :(
"Style is required! This model does not provide a default style. You should always prompt specific style. Or use a style LoRA. If you don't give the model style conditions, the model will give you the "ultimate AI style" that averaged every style because of dmd2. This is a "feature", not a bug. I don't like bake a strong style into the model, I prefer having choices."
You need to add some "@artist" tags.