Anima LLM LayerWise (FP8-FP32)
Use this in place of Qwen3_06B it is loaded as CLIP
FP8 VERSION TENDS TO SMOOTH DETAIL COMPARED TO FP32 or BF16/FP16
Anima Artist and Character triggers fully working
Full FP32 Training - Adam not 8BIT
Uncensored LLM can be used for story or text generation
FP8 Model has best of both worlds with 50% weight reduction and FP32 blocks where it counts.
Comfy UI users --fp32-text-enc or --bf16-text-enc can be placed in your launch script to force BF16 or FP32 casting of the LLM
Note: Top rated images where used for prompts as I do not know Anima well.
Description
FAQ
Comments (14)
can you upload the fp32 or if you have fp16 also?
Yes
What this even is? Lora? base model?
It's a text encoder, which Anima uses as CLIP. @Felldude mistakenly set it as checkpoint.
It is loaded by comfy as a CLIP, it is technically an LLM and still works as such - However if set as CLIP or Text Encoder then Civitai does not correctly show what resource it is intended for
With Anima LLCM LayerWise FP8 the quality is sadly generally slightly worse compared with the standard Qwen 3 0.6B CLIP and the size is only about 11% smaller :/
Anima LLM LayerWise FP8 1MP
Qwen 3 0.6B 1MP
Anima LLM LayerWise FP8 2.4MP
Qwen 3 0.6B 2.4MP
https://huggingface.co/Qwen/Qwen3-0.6B-FP8 would likely have less errors if tested over 1k images seed to seed.
However if used as an LLM the results of that model are also not uncensored, this model is mostly uncensored although hallucinates when asked math questions sometime compared to base.
just tested this model, I am more satisfied with the result when it comes to simple comic.
Thanks
Thx for taking care of Anima. :-)
I've just tested it while in a middle of a generation session and ... It doesn't tokenize "girl" like Pony or Illustrious, but like Flux. i.e. the 6-8 years old girl. This lead to immediate deletion of generated image.
I'll probably need a specific gen session to fully test this beast. ;)
I just wanted to warn other people about the "girl" oups.
This LLM is trained on QWEN chat template and not this dataset. However it likely has some answer to that if you look at when male female, man woman is used vs the WD14 style tags.
I did not have any issues with 1girl in anima base 3 preview - The images mostly came back large breasted anime females
Cant say that this TE is actually a better alternative. Its definitely not worse, its just... different?
Based on my limited testing (on the same seed), the images come out a bit more detailed sometimes.
NL prompt understanding is roughly the same, tag prompts understanding feels a bit worse, nsfw prompts (both nl and nsfw) - kinda the same in terms of expressivity and concept understanding.
Based on my experience - TE was never the problem with the censorship of the diffusion model. Even tho Anima is completely uncesored as well from the beginning.
What exactly is this model, I wonder? Is this just a finetuned qwen3 0.6B?
QWEN never released FP32 variant of 0.6B - as this is a full finetune it likely would train better assuming you tokenized in FP32
It also will return uncensored LLM responses so it could be loaded in comfy and used in think mode to enhance prompts (Not directly)