First thing first, I want to say a thank you to everyone who's been trying out, sharing, and supporting the model thus far. In less than 12 hours after release, Anima-2.9B has already received native support on major platforms including ComfyUI and Forge-Neo.
The model is also available to download from Civitai, where you can share your image and what people had managed to do even with this Preview version, like the RDBT Distilled Turbo LoRA for instance.
Anyway, LoRA training is now supported via my training GUI and you can train + load LoRAs natively in both ComfyUI and Forge-neo. A fork of regular sd-scripts is provided here, and a PR has also been created.
I've been also listening to your valuable feedback, and where the model can be improved even further in the next iterations, and I will explain what this version is and isn't, and what it aims to be in the future.
What v1 preview actually is
Regarding this version, perhaps it's more appropriate to name it v0.1, I guess? The released version was trained using Muon optimizer, on approximately 2.5 Epochs, but the true number is closer to 5 due to extensive use dynamic repeats on both new and old characters, with the dataset focus on post September 2025.
This version's main aim is to be more knowledgeable and be the most up-to-date anime model at release, v1 preview is not a "full" finetune (the whole original weights are frozen), and many pros and cons from the original Anima also carry over. By itself, v1 has its own strengths and weaknesses as well. It's not a model trained for aesthetic or with any RLHF. It's a model with a slight bias toward modern East-Asian anime style illustration, this is in fact intended.
It's also soon and easy to realize that 1.7M is not a very large number of samples (it's a decision I have to make at the time based on time and money constraint), and the new expanded layers has a lot of room for much, much more information. In other words, a lot more samples are needed, and that's why there is a 10M (pretraining non-anime focus) samples floating around. Unfortunately it's not a cheap or fast task to accomplish, and for that reason, your supports are greatly appreciated. Even without actual monetary support, it can still be achieved, just won't be quick, nor reaching the full vision/potential that I had for the final model.
Prompting guide
Finally, if you're struggling to prompt your desired results, here are several very important points for consideration:
- Characters should (think "must" in this case) be follow by their series/copyrights, (think of these like anchors, they always tag along) follow by their appearances (the more the better). Simple or very short prompt won't do as well.
- Don't use underscore except for score tags.
- Several metadata tags are very good to keep, I always recommend include highres and absurdres, following by the year tags (this has very strong influence on the generated image), score tags may not needed but you can still use them. You can throw away garbage such as "raytracing" or "4k" and "8k", these has never done anything and will just poison your output.
- Always recommend using artist tags, same as Anima-base, and you can mix them as I often do with proper prompt weighting, but don't expect it to be the same as sdxl.
- Prompt weighting and negative prompt are very important as well, this is something very easy to be underutilized.
- Prompting the background is also important if you want it to be more dynamic. Additionally, use keywords such as "cinematic composition" and "dynamic angle" can improve your image significantly.
Thank you once again, I will await your feedback.
You may need to install ComfyUI-Anima-2.9B to the custom node folder. Plug and play, there is no workflow node needed. Sometimes may not work with other custom nodes
Overview
Anima-2.9B is a fine-tune and layer-expansion of circlestone-labs/Anima. The base Anima model targets anime, illustration, and non-photorealistic art; this release continues training on that foundation with an expanded architecture. The model is trained on an additional 1.7M anime/illustration samples, with knowledge cutoff in July 2026, making Anima-2.9B one of the most up-to-date anime/illustration model at release.
Versions
Anima-2.9B-preview-v1: initial release
Training/Dataset
Trained using Muon optimizer on a 8x 5080s cluster, with earlier steps trained locally on my PC
As of preview v1, only new layers have been trained, with roughly 70% of the compute spent on 1024px
Knowledge cutoff in July 2026, training data included both new and old samples prior to September 2025
Mixed captioning, including both tags and natural languages, using a mix of Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, and Claude Sonnet 5
NO scoring
Architecture
Transformer depth expansion: expanded from 28 transformers layers to 40, growing the model to ~2.9B parameters. Each new layer is added by deep-copying its neighboring layer's weights, using interleaved insertion with zeroed-out output projections, making the new model functionally identical to Anima-base at initialization.
Prompting tips :
Follow Anima prompting tips: quality tags, year/period tags, @artist tags, character count (1girl, 1boy), character tags (follow Danbooru and Gelbooru tags), series/copyrights, base appearance.
Character name/tags should be follow with series/copyrights tags or else the model might confuse.
For multi-character images, attribute the character and names with their respective tags/appearance.
The model does improve the base art style slightly, but I'd still recommend using artist tags.
The dataset does not include score in its captions, however, you can still use them. Also HIGHLY recommend using year tag.
(IMPORTANT) THE MORE DETAILED THE PROMPT, THE BETTER, short prompt will often generate a bland simple background, and may not able to produce the desire results
Generation (Recommendation)
Sampler: Euler/Res-multistep/Er-sde
Scheduler: sgm-uniform/beta/beta57/linear-quadratic
Resolution: 812x1216, 1152x1536, 1536x1536
Steps: 28-50
CFG: 3.5-5
My personal usage are euler + sgm-uniform, which has a good balance between composition and fine details. Additionally res-multistep + linear-quadratic spend more time at high noise steps, which does lead to visibly better composition. My recommendation for the highest quality is 50 steps, CFG 3.5. You may need to use slightly lower CFG than regular Anima. Experiment yourself!
License
Model weights are released under the CircleStone Labs Non-Commercial License, falling under derivative model category.
Not affiliated with CircleStone Labs
Acknowledgements
Built on nvidia/Cosmos-Predict2-2B-Text2Image and circlestone-labs/Anima.
LLaMA Pro: Progressive LLaMA with Block Expansion.
Training infrastructure built on sd-scripts.
Description
Initial release
FAQ
Comments (86)
Is it just no scoring on the new data or on the old data as well?
no scoring on new data, but you can still use them
does this has updated character ?
yes, I put some of them on the post
I don't know why the output is always snowflake shaped images
what do you mean by that?
@gazingstars321 Image generation failed, but Anima1.0 is capable of generating images
I am using qwen_3_06b_base.safetensors and qwen_image_vae.safetensors
@280302610211 you need the custom node if you're using comfyui
if you are using latest neo 28.1, you could modify "detection.py" at line 196 with https://github.com/Haoming02/sd-webui-forge-classic/commit/d1b4fdfe37040dcce25b0bfd0cb34dc0380d87cf#diff-4333b16cc1dd2b06a21479514dad5fc4995823134dfaed009656a49737bef39dR196 or wait for the next release.
@riel_mayer what a weird non noob freindly answer.
Here, https://github.com/Haoming02/sd-webui-forge-classic/blob/d1b4fdfe37040dcce25b0bfd0cb34dc0380d87cf/modules_forge/packages/huggingface_guess/detection.py
Download this (top right, raw file) and replace it with your exiting detection.py
@starchaser92 ikr trying to help is super sketchy, shame on me. gold star for your "non" weird answer
giving this a like and comeback later once i tested on my poor 3050
I've downloaded the custom node into the Custom Nodes folder, and reset Comfy, but the loader doesn't do anything.
It only renders blank images when using this model. :/
Is your comfyui updated to the latest commit? If so, try not using the custom node instead
厚礼蟹！This is absolutely fantastic! Perfect, just perfect!
Amazing model, thank you so much for creating this! It is so refreshing to see updated character knowledge with Anima. I updated my article comparing different knowledge cutoffs and can confirm that this model here indeed has a knowledge cutoff of July 2026, which makes it the most up-to-date Anime model at the moment. In comparison: Anima 1.0 and v1.1 both have a knowledge cutoff of September 2025 ^^
https://civitai.red/articles/14008/character-knowledge-cutoff-comparison
I am confused, does this not go into diffusion models folder?
goes to .../ComfyUI/models/diffusion_models/Anima-2.9B-preview-v1.safetensors
Kind of a stupid question here, but the guide recommends using 'year tags'. What does that mean? I understand that a year is like 1995, but under what context should I use it and for what purpose?
When an image in a training data was made, you can use it to replicate old generic image styles, but I don't know how reliable it is.
@who_is_civet Ah interesting. The page says "HIGHLY RECOMMENDED" but your description makes it sound pretty optional.
Newer years tagged images have better quality in general, so using "year 2025" in the prompt works just like a quality tag. But you can also try to make use of it targeting a specific year for a style or something.
@who_is_civet to expand on this if you put "year 2010" images will look more similar to the "common style" of that time period. It's basically a style filter without artist tags--an amalgamation of all art trained from that year. As a practical example, if you have an artist whose style has shifted substantially over the years, you can use the year tags to suss out their original/later styles specifically.
@sansmia I don't get how they caption those because Danbooru doesn't allow using year tags to indicate when an image/video was made (unlike e6, which does allow that)
@who_is_civet Pretty sure it's the upload date.
Any plans for a turbo lora?
is there a valid artist list for this model ?
Updated tag list where artist tags are prefixed with @
https://github.com/BetaDoggo/danbooru-tag-list/releases/download/Model-Tags/anima-2.9B-preview-V1.csv
@qzd Thanks
Great work, loving it!
Shame that I need now to re-train some of the LORAs ;/
can I still use Anima loras and controlnet?
Great model. There's a clear prompt comprehension upgrade compared to base Anima. Also, the same prompt tends to have more variety. Looking forward to your progress.
Pretty interesting model, definitely excited to see how it'll come out in the final release. Hoping that it will make better backgrounds than the current Anima can at least, I wasn't too impressed with Aesthetic 1.1 or 1.0b in terms of better details or backgrounds unfortunately.
I do have one question though, what was the reason for adding more layers to Anima? Did you try finetuning Anima base normally at first then decided that adding more layers would be beneficial? Since this model is slower than Anima base in both inference and training I'm wondering what led you to make this decision over just finetuning the model normally.
This model is an amazing upgrade to Anima! But the fact that no old lora works on it is such a punch in the butt T_T
https://github.com/R0smontis/ComfyUI-Anima-28to40-Lora-Converter
convert old anima lora with this node
Is there anything for forge-neo to do the same?
Really saved me, thanks pal
Apparently forge-neo does it automatically for anyone wondering.
Ohh I just replied to someone about this. Thanks a lot!
How to use it ? How to use it in comfyui ?
Download the custom node, or update comfyui to stable version 0.33.1
Great! Thanks for sharing your model!
It works with Forge Neo?
Yes
@gazingstars321, When I try to generate something with Euler SGM Uniform, image contains only pixelated squares like colorful noise. What do I do wrong?
@MarkinZzZ you need the latest update
@gazingstars321, you mean 2.28.1? I think I got it. I used git pull to update, but webui user is still showing ver 2.28. It's supposed to be like this or I did something wrong?
@MarkinZzZ 2.28.1 is stable version from 2 weeks ago, you need to git pull the latest commit instead
@gazingstars321, I don't really into github commands. What command do I need to use for it?
@MarkinZzZ https://chatgpt.com/
int8convrot: https://civitai.red/models/2855769/anima-29b-int8-convrot
16 steps turbo and quality booster LoRA: https://civitai.red/models/2852092/rdbt-or-anima29b
Can't it incorporate the styles of multiple artists? When I @ multiple artists' tags, they don't seem to work.
you can mix them as I often do with proper prompt weighting, but don't expect it to be the same as sdxl
First, I’d like to apologize to the creator of this model for not looking into this powerful model carefully enough before. Thank you for your contribution.
⚠️ IMPORTANT: You need a Forge Neo version with a commit of at least 81e1dfa.
Here’s what I found about why LoRAs trained on anima_baseV10 don’t work properly with this model:
1. Block mismatch: Most LoRAs for anima_baseV10 were trained with 28 blocks, while Anima 2.9B uses 40 blocks. Without an intermediate conversion step, Anima 2.9B cannot properly interpret LoRAs made for anima_baseV10.
2. Forge Neo already fixes this: This issue has been addressed by Haoming02, the current maintainer of Forge Neo. No LoRA retraining or manual modification is required — the conversion is already built into Forge Neo.
⚠️ IMPORTANT: Make sure your Forge Neo commit is 81e1dfa or newer.
Older versions will not automatically convert the LoRA, meaning Anima 2.9B will not be able to properly understand LoRAs trained on anima_baseV10.
In short:
28-block anima_baseV10 LoRA → Forge Neo conversion → 40-block Anima 2.9B
So if your anima_baseV10 LoRAs weren’t working, the LoRAs themselves may not be broken — you may simply be using an outdated Forge Neo version.
https://postimg.cc/Cd2NyMF1
This is what I tested. None of the LoRAs I trained with anima_baseV10 work on this model.
An IQ too high?
@razdel19 I’m just expressing what I tried and the problems I ran into. Is there something wrong with your smooth, flat little brain or what? 🙄
Ok calm down. This is on purpose. This is what the model is all about
@tranthai >"I’m just expressing what I tried and the problems I ran into"
No, you wrote "This model is completely broken.".
We can see who has the small brain here.
@gazingstars321 So how can I get Forge Neo + your model to produce the correct results?
@nogo Oh, and you weren’t helpful at all :> So I guess you’re on the same level as my tiny little brain after all. What’s the difference?
@tranthai if by "correct results" you mean work with loras for the original: you don't. Description says pretty clearly this is a very large fine tune/overhaul, there's no reason to expect base loras to work. You literally said in your first comment "You could basically consider this a completely new model" ... which was both correct and does not mean "broken".
@tranthai retrain the lora
It actually does work, ya dingus.
Here I tried in a hurry and it worked immediately: https://civitai.red/images/139701127
@nogo Is this question really too simple for your simple-minded intellect? I’m sorry for overestimating you.
Don’t be sad, though. I’ve figured out the reason already, while all you could do was tell me that it works without even knowing why. 🙃
Still, thanks for the simplicity of your brain. It helped me figure out the cause and fix it, even though it was honestly pretty frustrating.
@tranthai Go back to kindergarden, kid. You are a lost case
@nogo Poor you. Couldn’t even handle a question this basic. And after I’ve already explained exactly why, you’re still like this? I honestly feel sorry for you. If you can’t understand what the question is actually asking, then while I’m thinking of going back to kindergarten, perhaps you should go all the way back to tadpole and start practicing how to communicate properly. Might help with the limited intellect situation.
@tranthai thank you for your help, ive updated forge neo and my loras are working again. do you think i should train future lora using anima-2.9b over base? does higher block also help with learning? i'm a newbie im sorry
@airuru If you don't have problems retraining later, sure. If you want a more stable version then use the normal anima. I'm retraining some loras for this version as a test
@sovietlat Got it! Im also retraining one lora just to see if there is any difference but for simplicity ill use the base to train future loras
why do none of your images have metadata
are they not directly from the model
Pull them into comfyui workflow, or download and inspect the metadata directly through notepad
It may be my luck, but I think this was trained on doujinshi sites or something like that, speech bubbles are a lot more present. They go away using the negative prompt tho
I made an extension to make the original anima lllite controlnet model compatible with anima 2.9b
https://github.com/SakanakoChan/ComfyUI-Anima-LLLite-BlockMap/
Mostly it has better output than default Anima, better anatomy, better style recognition. I wonder will finetunues make it even better
I know this requires much more compute, but it would be much better if you switched and retrained LLMAdapter to hook all TE layers as Krea 2 did. I think this would give a huge boost to prompt adherence.
Anyway, thanks for your hard work!
Llm adapter are trained. Prompt adherence can be improved by training with much larger dataset
@gazingstars321 Hi, thanks for your hard work! Just for curiosity. I wonder what's your thought about the Mage-Flow (main repo 404) model from Microsoft. It's a powerful yet extremely cheap model (2x params, but 2x faster than Cosmos cp2, also 2x faster to train, back to SDXL level), with a powerful TE, and under MIT license. It has built-in "controlnet" support via pose keypoints, canny, etc. already. Makes pose control much more easier to finetune and have better quality. It has so much potential.
i did find : res-multistep + linear-quadratic with cfg of 3.5 much better for accurate complex poses than euler + sgm-uniform. overall best anima model out there so far.
styles are weaker, even with styles that are already weak with the base model. i also get worse hands. using a1111 forge neo with recommended settings. not using any loRAs.
maybe i did something wrong, not sure.
Its absolutely amazing for newer data and I love it. However, styles are much weaker, and if you combine multiple artist tags, it results in the image looking like neither artist and seems to take on a more default style. Really sad as I havent found a fix for it, if it wasnt an issue I'd honestly abandon anima base.
Interesting. I get mixed results with using style loras, some of them (I think) work, others don't.
OH! Someone shared a comfyui Lora converter in the comments below. I gotta try that out.
I think its more because this model is still in preview stage... If the training continues prob this issue would be fixed
I also encountered this problem. A single artist tag works, but when there are multiple tags, it doesn't work anymore.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.