v1.0: I think it's time to release v1.0. I've migrated to the krea2 model (never imagined that krea2 would be my next choice...). Fine-tuning the anima model is really fun.
I also uploaded int8 version for ComfyUI hardware int8 (labeled as fp8).
About 8/16 step distillation:
8-step has higher stability. Easy to use and prompt.
16-step (same as previous old versions) has higher diversity and more dynamic details, less sloppy but less stable, more small errors. (I still prefer this one because anime/art model needs a little bit creative and chaotic.)
Comparison images are uploaded as 16-step version showcase images. You can feel the "complexity" of 16-step is noticeable higher than 8-step.
FYI: More distillation = higher stability = lower steps = lower diversity/complexity = more slop-ish. For reference, official anima turbo is 4-step distillation, means highest stability and lowest diversity.
Interesting old version:
b1 v0.39.b: compared to v1.0 16-step, more creative, more dynamic details, but more logical errors. Good for creative, abstract effect, etc.
b1 v0.37.1: 2-stage finetuned, highest aesthetic among all versions, but due to small dataset it has bias. Also has several built-in quality/style trigger words, see update log "b1 v0.32".
EDIT: Why Krea2?
It knows everything. Default quality is off the roof already. (Of course, it's a large 12B commercial model.)
It knows anime/art design, even knows popular characters. (Means it has been trained on a much more creative art dataset. Previous open source models didn't, they mainly for photorealistic)
Official turbo model, 8-step. (FYI: Turbo model is as fast as non-turbo 20-step Anima.)
If you want anime art. All you need is a style/character LoRA.
Life is much easier...
RDBT [Anima]
This is a finetuned model with 10k high aesthetic images paired with natural language captions from LLM. Then distilled to further improve quality and stability. Dataset does not contain any shiny plastic glossy AI image.
It's not overfitted and doesn't have a default style. I use it as a clean starting point to stack more style LoRAs. I can stack whatever I want and get exactly what I stacked.
See this page for update log.
For advanced users: The RDBT model is trained as LoRA natively. See this page for original LoRA.
This model is based on:
prefix with ym: AnimaYume (hf link) (civitai link).
prefix with b,p: Anima pretrained (hf link)
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
If someone is selling this model as their own, I'm happy to list them here so everyone knows.
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".
Usage:
Settings:
CFG scale: 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 recommended
Steps: 16+
Prompt:
Always specify style, or use a style LoRA. Otherwise, you will get random/mixed style. This is a feature, not a bug. This model does not provide overfitted default style.
Quality tags:
It's recommended to omit ALL quality tags. The fine-tuning dataset has higher quality than "masterpiece". Thus they don't have noticeable effects. Omitting those redundant tokens allows LLM to pay more attention on other words
Training settings
All captions are NL from Google Gemini.
Optimizer: adamw, constant lr 0.00002, weight decay 0.1, batch size 16.
LoRA rank/alpha 24.
Timesteps shift 3.
Block 0-2 and adaln linear layers are skipped.
Description
FAQ
Comments (39)
We might finally have the "Illustrious killer"
Good to see NoobAI's approach of focusing on artist names.
Model is still in training, so results are can suffer from bad anatomy or inconsistency. Also artist names are not an addition like with Illustrious, but almost a requirement to get good results. Use them like @{artist name}.
This model has zero involvement with the NoobAI team!
@NamelessKing2 I know, if I am not mistaken, it is one guy. I am just glad that shizo tagging of PonyXL is not standard
As for now it just feels like much worse, but easier to use and faster newbie image.
it's not bad but a bit disapointing with output quality
Prompt adherence is just insane!
Getting great results ♪(^∇^*). (it's uncensored btw, understands a lot of concepts, better than lumina)
Hope this model gonna hype up, cause it is a potential Illustrious killer.
On my 3080 (12 gb), 30 steps it takes 22 sec per image generation
Uhm... where is the patch? i m missing something?
it is for old GPUs like turing architecture, if have 30 series or newer than you not need to worry about it.
I tried it, and placed the file in this directory: ComfyUI\custom_nodes\anina_fp16_patch.py
@krigeta I have an rtx 2060...so yea
this patch is good for amd gpu's as well
The model is a beast, fast, real good quality even on base res gens without hires, very good prompt adherence, more artist and character knowledge than Illus and nsfw capable with regular language and danbooru tags, and this only the preview model lol.
Thank you. Where do I put the patch file?
I tried it, and placed the file in this directory: ComfyUI\custom_nodes\anina_fp16_patch.py
lora training and controlnet and we got the best anime model right here of 2026.
I dont understand, what are those files? Is circlestone labs released fp16 model?
Hello, can you also add fp16 patch to z image base?
read on comfyui github he did a patch for z-image but can't find it now
@xpnrt that patch no longer works.
@3059714361 no, z image base is way more f**ked up than turbo. Seems they continue training the model, a lot. I guess the "base" mode we see right here, is not the real base model of turbo.
Holy this model is insanely good also MOSTY follow your prompt
The model is great! :3 can't waut to see others use it!
Can someone please explain why I'm getting a mismatch with the suggested text encoder? Despite using Qwen_3_06B_base, I still get "Error(s) in loading state_dict for Qwen3_4B: size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 1024]) from checkpoint, the shape in the current model is torch.Size([151936, 2560])."
I've tried both anima_preview and animaFP8_preview, as well as changing the various load clip types (Stable Diffusion, Cosmos, Qwen Image). Thanks in advance for any response.
Update your comfyui?
@ZootAllures9111 Thanks for the reply.
After updating comfyui, everything worked as it should 🤷.
FYI, the FP16 patch will not be faster on 1650 / 1660 series Turing GPUS, they don't have proper FP16 support.
Thank you so much for creating this patch! However, I encountered an issue during sampling: expected mat1 and mat2 to have the same dtype, but got: float != struct c10::BFloat16. > Does this patch only support FP8? I am currently using the original version of Anima. If that’s the case, I will switch to the FP8 version instead. 🥹💕
I initially encountered a similar issue; simply removing the UNET settings in FP16 resolved it.
Perhaps it conflicted with a patch, so restoring the default settings might be necessary.
@dawn66666666 Thank you for your reply! How should I go about deleting the 'fp16 unet setting'? I'm so sorry, I’m not very tech-savvy with this. 🥹
@shigjfg I apologize, I'm not an English user, and I'm not sure if the following translation will be accurate. In my ComfyUI launcher, there's an advanced option called "Calculation Precision Settings," where you can set "fp16, bf16, fp8, or determined by ComfyUI."
@dawn66666666 okay,thank you very much for your answer.❤️
The patch made my anima gens 6x faster, thx man
Thanks, this greatly improved the speed of my Tesla V100; it now takes about 35 seconds.
It used to take 130 seconds.
It's a f***ing amazing model! Know a lot of characters, A LOT of styles/authors, follow the prompt, very fast and lightweight....
Well, my job is done here. Why should I retrain all my styles for ZIT/Klein when this model perfectly replicate almost all of them out of the box, and it's not even a final version?
12/10 for model, and HUGE respect for a company that produced it.
You said it, it's killing creators and you're happy about it 😂🤣
@GlowingGuardianGirl There are still many things that this model lacks, so there is room for improvement. Because booru-dataset contains... mostly very specific artists.
But yes, the ideal model is one that can do everything out of the box and does not require additional training =)
Dumb question,where should I put the patch to apply it?
# How to use:
# Put this file in the ComfyUI "custom_nodes" dir.
# Use "ModelComputeDtype" node and set dtype to "fp16".
# To disable the patch, remove the file, or rename the ".py" suffix, to something
# like ".disable", whatever.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.