RDBT [Anima]
This is a general finetuned + distilled model.
Better quality, better prompt adherence.
Dataset contains ~10k handpicked images with accurate NL captions from LLM, body/hands anatomy. Does not contain any shiny plastic glossy AI image.
Some cover images look ridiculous; they're just for demonstration purposes for corner cases.
I use this model as a clean starting point to stack more LoRAs.
See this page for update log and version info.
For advanced users: RDBT model is trained as LoRA natively. See this page for original LoRA.
For those who don't want to stack style LoRAs and is looking for a out-of-the-box ckpt: RDBT | Anime, which is based on RDBT finetuned base, but optimized for high quality digital art.
Links to base model:
prefix with ym: AnimaYume (hf link) (civitai link).
prefix with b (base), p (preview): Anima pretrained (hf link)
Usage:
Settings:
CFG: 1~3. This 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. You don't need those. 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.
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 (14)
I believe you said some time ago, that you didn't feel like training with your full dataset because of how the circlestone labs license is structured. Is v1.0 of RDBT now a full dataset run or still limited in this regard?
Still limited, but that doesn't matter. Current dataset is handpicked. My larger dataset is fully automated and the avg quality tbh isn't good.
16 step V1 is better than 8 step V1. I do find that I have to either add or be a bit more specific with my prompts than before compared to b1V039B.
Edit: Perhaps I should be more specific in my comparison. For V1 16 step I would have to be more specific in regards to lighting and placement of furniture compared to b1V039B. For V1 16 step, I would have to say "Bed on right side of view" or something like that or else I would get what I would describe as illustrious style small and disfigured beds on either side of the image. Not the worst thing in the world, just requires more wording to get what I want in some scenarios but others kinda get wrecked when it comes to two people and hands interacting in very specific scenarios.
Edit 2: Noticing that prompt adherence seems to vary from prompt to prompt and is seemingly affected by both sampler and schedule type. Not sure why this is unless I just never noticed before on other anima checkpoints. Some prompts would do better with euler/euler a and others would do better on ER SDE alongside scheduler like simple/beta/normal
any idea how to merge int8 models in comfyui? I normally merge rdbt with anima base, but the int8 variants don't seem to work that way
I don't think ComfyUI can merge quantized model, also this is not recommended due to quantization loss, normally you merge bf16 models and quantize it .
You can use the int8 anima base model, and add rdbt LoRA.
You can merge bf16 model in ComfyUI and use this tool to quantize
https://github.com/silveroxides/convert_to_quant
There are some custom nodes that supposes quantization in ComfyUI, but kinda buggy.
Using the Anima Turbo Lora v0.2 on latest model (fp8 version) produces images nearly identical in quality as to just using it normally, so very excellent in speeding things up!
does latest rdbt anima version has any external requirements? It wont start, giving no error if just plug instead of anima base
int8 version needs comfyui v0.27
Thank you for your amazing work. Your RDBT model is truly magical. No matter how bad my training data is and how badly my character LoRA perform on the base anima model, the rdbt LoRA can correct it, without bias, and much faster.
Hope you have fun with krea2. It's too big for many people's hardware.😭
Edit: int4 now in ComfyUI!
Krea 2 turbo model (cfg 1, 8 step) is as fast as non-turbo ~20-step Anima. Which isn't slow. int8 version is 13GiB, and comfyui is adding int4 support.
Thanks for the info!
waiting for the int4 support, impatiently.
Model is so cool! In my opinion, the best of the Anima type model. I would also like to add something like Arknights Endfield to the dataset, but this is purely a wish, and everything is already good.
Why can't I merge RDBT with other Anima models? It says "Keys Mismatch".