RDBT [Anima]
A general finetuned model, better overall quality, better prompt adherence, less artifacts and errors.
Dataset contains ~10k images, handpicked by professional amateur artists, captioned by latest LLM. Accurate body/hands anatomy. Complete background.
Some cover images look ridiculous; they're just for demonstration purposes for corner cases.
Update log and version info: link.
Original LoRA: link. For advanced users: RDBT model is trained as LoRA natively. You don't need to download this full ckpt.
My distillation LoRA for RDBT: link. I won't recommend the official Turbo model because it's very sloppy and overwrites style LoRA. I have my own.
Links to base model:
prefix with ym: AnimaYume (hf link) (civitai link).
prefix with b (base), p (preview): Anima pretrained (hf link)
For those who don't want to stack style LoRAs and is looking for a out-of-the-box ckpt with high aesthetic style: RDBT | Anime, based on RDBT, and optimized for high quality digital art.
Usage:
Settings:
CFG: 1~3. By default, model has been distilled. You can disable CFG (CFG 1) and run the model 2x faster. Cover images are without CFG for demonstration. "RenormCFG" node is highly recommended if CFG is enabled (CFG > 1), set "renorm_cfg" value to 1.1.
Steps: 16+
Sampler: Euler (best diversity), Euler a/er_sde etc. (better stability)
Res: 1MP
Prompt:
Always specify @style in prompt, or use a style LoRA. Otherwise, you will get random/mixed style.
Quality tags:
Omit ALL quality tags. The fine-tuning dataset has higher quality than "masterpiece". Thus quality tags don't have effects. Omitting those redundant tokens allows LLM to pay more attention on other words.
FAQ:
Why I won't bake a default style:
If a model has a "default" style (ignores prompt and is always active). We call it "overfitted". Technically, it's not a feature, it's a bug. This means the training dataset is not big enough, and the model has lost its ability to generate other styles. I know some default styles look very cool, but this is not my goal.
The goal of RDBT model is to "perfectly" reproduce styles. As some of you know, I use this model to aid in drawing. I need uniqueness (diversity) and consistency in style.
This model has excellent compatibility. You can stack any type of Lora on it. After all, it's you who decide what style you want, not me.
RDBT can't remove nor resist overwhelming Al slop effects
Too long, see: link.
Training data and why it works
AI model is a garbage in, garbage out. What we want are production wallpaper level images. While the Danbooru dataset is massive, high-quality data is extremely scarce. 99.99% imgs are "incomplete" because artists are lazy drawing is hard. Leaving rough outlines, incomplete backgrounds, etc. Therefore the model will also generates "incomplete" images. I "fixed" this by training the model with 100% "complete" imgs.
Sharing merges using this model is not allowed.
This mode is a free and it will always be free. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Known model thieves: NukeA.I (selling this model behind paywall on tensorart).
Because he is such a jerk I wrote a story about it. Also contains a guide for trainers about "how to bake special trigger word into your model".
Description
FAQ
Comments (7)
Whoa, this turbo is actually quite amazing what the hell? Much better than normal Anima Turbo in terms of quality and just as fast, this is a god send.
I know v1.0 has too many versions and are confusing. A brief clarifying:
All v1.0 shared the same finetuned base model, but later distilled with different settings.
16-step: is the most creative one. More diverse, dynamic.
8-step: used standard distillation settings that a "turbo" would use. Needs less steps, more stable, but less diverse. (This one probably can be deprecated in favor of the official turbo model.)
Those two versions above are homemade. I don't have huge amounts of resources, I can't take care 20k built-in artist style tags, they're too many. Thus built-in artist tags are not distilled. Without CFG, they have weaker effects. (This is not a issue for new styles from external LoRA).
E.g. 1, 2, (16-step is more dynamic than 8-step) 3 (built-in artist style is weaker)
turbo: this one merged new released official anima turbo model. This should be similar to the "8-step", but has better built-in styles. E.g. 3 (built-in artist style is more accurate),
TBH my favourite combo is still v0.39 (the last one before 1.0) + the new Anima Turbo model as base. It has considerably more creativity and style adherence than all 1.0 ones in my testings.
My favorite combo is v1.0 16-step + new anima turbo (extracted lora). 0.5:0.5.
Official turbo really bad at lighting, always locked to mid brightness. v1.0 16-step can go really dark and cinematic.
In LoRA usage, the 16-step seems to be on par with Turbo; personally, I even lean toward the 16-step model.