RDBT [Anima]
This is a general finetuned model, based on pretrained Anima.
Better overall quality, better prompt adherence, less artifacts and errors, and highly creative.
Dataset contains ~10k handpicked images. NL captions from LLM. Accurate body/hands anatomy. Does not contain any shiny plastic glossy AI image.
It's a base model for every style, not a model that has a default style. I use the model as a clean starting point to stack more LoRAs. I can get the style exactly as it should be.
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 made 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. This is a feature, not a bug. This model does NOT have overfitted default style (which ignores prompt and is always active).
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:
About my distillation:
Compared to the official turbo model, my distillation has very high diversity (seed, style diversity, etc.), and is more compatible with style LoRAs (it won't add a quality filter which changes/overwrites the style drastically).
I didn't name it "turbo" because it needs 16 steps.
Why not bake a default style?
If a model has a default style, this means whether you prompted it or not, the style is always active. This is called "overfitted", this not a feature, it's a bug.
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).
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 (11)
ooo more steps approved model (0.32b) gota see that.
0.32 is pretty good at text (comparable success rate to v0.24 (maybe bit worse than it) = much better than v0.25 to 0.29) and can be mostly convinced to follow styles (get lost 3d flash-bang), also does compositions well (still has stroke at times, but its better/ as good as 0.24), just at times does not understand abstract things (for some reason very stubborn at not generating image that does not distinguish between floor and wall) but that minor issue :) +10
the latest version definitely feels like an improvement over previous versions. Styles work a lot better now, better prompt understanding, less 3D/generic slop bias, more dynamic poses and composition,backgrounds can get pretty detailed as well. I'd say it's a big win.
Boss, now that the 1.0 version is out, I'll leave here my request for you to do your magic in a version that doesn't degrade diversity (perfectly aware of what that implies for generation time/stability). IMO the ideal choices for your loras would be a stability focused distill, like the ones you've been cooking so far, for quick and pretty anime shots, and a diversity focused one, when you actually want it to be a suitable replacement for Illustrious.
I give up doing step distillation. It's too much for me.
Love 0.32.b
v0.32 is such a banger.
I'm dropping step distillation. 1) my cheap distillation really kills the quality. yes, 4-step works but the quality is sh*tty worse than extracted cosmos lora from a very very far away model. Then I changed it to 12-step distillation, looks better but the problem is still there.
2) seems anima official has their plan to do step distillation (aka, turbo, 4/8-step model). They have the money and recourse and full dataset. I don't.
3) if you need higher stability or speed, you can stack the extracted cosmos lora or the anima-turbo, basically can achieve the same thing, probably even better
--
but this model will still be guidance distilled. So you can still use cfg 1 and run the model 2x faster and get similar output. I consider it as a free improvement. It adds a little bit stability, enough to fix many small errors like hands, and won't noticeably affect diversity, and is cheap to train.
damn 😢️
also screw model thieves
Thank for still making one of my favorite models x) ! it's totaly understable at this point after seing what they said and the liscence fee and all that to let them do it.