CivArchive
    Kirazuri (Anima) - v3.0 [anima-base-1]
    NSFW
    Preview 133377774
    Preview 133377779
    Preview 133377990
    Preview 133377968
    Preview 133377950
    Preview 133378813
    Preview 133377898
    Preview 133377928
    Preview 133416363
    Preview 133438844
    Preview 133380625
    Preview 133380628
    Preview 133378161
    Preview 133378641
    Preview 133377680
    Preview 133381026
    Preview 133378606
    Preview 133378584
    Preview 133380653
    Preview 133450150

    Kirazuri (Anima)

    Kirazuri (Anima) is a full fine-tune of the Anima Base v1.0 model by CircleStone Labs focused on several goals:

    • Learn new concepts/styles/characters past the base model dataset cutoff of 2025 September

    • Enhance the model aesthetic guided by manually applied quality, aesthetic, and style tagging

    • Improve rendering and understanding of fine-details through high-resolution training for 1024^2, 1280^2, and 1536^2 resolutions

    Version 4.0 (Latest)

    For in-depth details of training and tooling, see:

    Training Details Summary

    Trainer: diffusion-pipe

    Training device: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition

    Total training time: ~10 days

    Total samples seen(unbatched steps): ~3,000,000

    Training resolutions:

    • 512^2

    • 768^2

    • 1024^2

    • 1536^2

    Stage 1

    • Samples seen(unbatched steps): ~2,000,000

    • Training time: ~125 hrs

    • Learning Rate: 6e-6

    • Learning Rate Scheduler: Cosine

    • LLM Adaptor Learning Rate: 8e-7

    • Precision: Mixed BF16

    • Optimizer: AdamW8bit with Kahan Summation

    • Weight Decay: 0.01

    • Timestep Sampling Strategy: Logit-Normal

    • Training Resolutions: 512^2, 768^2, 1024^2

    Stage 2

    • Samples seen(unbatched steps): ~1,000,000

    • Training time: ~84 hrs

    • Learning Rate: 2e-6

    • Learning Rate Scheduler: Cosine

    • LLM Adaptor Learning Rate: 2e-7

    • Precision: Mixed BF16

    • Optimizer: AdamW8bit with Kahan Summation

    • Weight Decay: 0.01

    • Timestep Sampling Strategy: Logit-Normal

    • Training Resolutions: 512^2, 1024^2, 1536^2

    Additional Features

    • Masked Training

    • Tag Dropout: 30% with protected first 8 tags

    • Tag Shuffle: Applied to last unprotected tags

    • Natural Language: Short and Long Caption variants

    Changes from Kirazuri (Anima) v3.0

    • Dataset includes recently curated 2,450 images increasing total size from 42,608 to 45,058 images

    • Dataset cutoff now of 29/06/2026

    • Introduced Masked Training for images with simple backgrounds

    • Updated tags+caption variants structure

    Installing and running

    Workflow:

    Reference the anima base instructions. The model is natively supported in ComfyUI. The above image contains a workflow; you can open it in ComfyUI or drag-and-drop to get the workflow.

    Note: Most preview images on the model card additionally use the custom comfyui-prompt-control node for schedule prompting syntax to mix concepts i.e. [word1|word2]
    This custom node is entirely optional but required to exactly recreate the outputs in ComfyUI.

    The model files go in their respective folders inside your model directory:

    Quantizations

    Int8 supported with latest ComfyUI:

    Distillations

    4 Step CFG 1 Turbo LoRA distilled with DP-DMD - diversity-preserved few-step distillation:

    Generation Settings

    Trained in mixed resolutions for the majority of training, and finished with dedicated high resolution training.

    Previews are generated mostly at 1280^2 e.g. 1520x1040 or 1536^2 e.g. 1248x1824 resolutions.

    30-50 steps, CFG 4-5.

    Same samplers as recommended for the base model work, I like to use:

    • er_sde: the recommended default for 30-50 steps.

    • sa_solver_pece: can converge with good detail in 15-20 steps.

    [Optional] ComfyUI-Autocomplete-Plus prompt input assistance

    An optional file danbooru_tags_kirazuri_4.txt is included to provide prompt tag autocomplete.

    This file contains metadata that is derived from public sources for prompt assistance only, and is intended to be used with the ComfyUI-Autocomplete-Plus extension.

    Rename the file to danbooru_tags_kirazuri_4.csv and place it in your ComfyUI/custom_nodes/comfyui-autocomplete-plus/data directory.

    You should see suggestions when writing comma separated tags in ComfyUI:

    Prompting

    Like the base model, this model is trained on booru-style tags, natural language captions, and combinations of tags and captions.

    Tag order

    [quality/meta/safety tags] [character] [series] [artist] [1girl/1boy/1other etc] [general tags]

    Mostly the same order as the base model, only the [1girl/1boy/other etc] groups position is towards the end in this models dataset.

    [quality/meta/safety tags] [character] [series] [artist] tag groups are also not shuffled, so their order may have some influence on generations.

    Quality and Aesthetic tags

    Human score based: masterpiece, best quality, very aesthetic, aesthetic

    The very aesthetic and aesthetic tags are where this model diverges from the base, with the intent these can be used to guide the model toward a different aesthetic - a kind of house model bias.

    Meta tags

    absurdres, official art, etc

    Styles

    painterly, chiaroscuro, ligne claire, flat color, no lineart, blending, etc

    traditional media, oil painting \(medium\), watercolor \(medium\), etc

    Known Limitations & Issues:

    Some concept bleeding and instability is noticeable when using short prompts, especially tag-only prompts.

    Longer tag strings and natural language prompts describing the image in detail should help with this.

    This reflects how the model was trained with a combination of natural language and tags.

    Recognitions

    • Thanks to CircleStone Labs for the Anima base models.

    • Thanks to tdrussell of CircleStone Labs for the diffusion-pipe trainer.

    • Thanks to bluvoll for support using their fork of diffusion-pipe.

    • Thanks to narugo1992 and the deepghs team for open-sourcing various training sets, image processing tools, and models.

    License

    This model is released under the same license as the base model.

    See the base model for details of the CircleStone Labs Non-Commercial License.

    Built on NVIDIA Cosmos

    FAQ

    Comments (42)

    HysocsJun 11, 2026· 7 reactions
    CivitAI

    Thank you for providing training info, we need more of this in the space. very good model

    Dewal76Jun 11, 2026
    CivitAI

    did you rent the NVIDIA RTX PRO 6000 or you bought it?

    motimalu
    Author
    Jun 12, 2026

    The model was trained locally.

    Seii1Jun 11, 2026
    CivitAI

    it has updated character ?

    motimalu
    Author
    Jun 12, 2026

    Yes it does, the dataset is manually curated so not all characters you may be interested in up to the dataset cutoff date are included however.
    You can check the "[Optional] ComfyUI-Autocomplete-Plus prompt input assistance" section on the model card for details about a plugin that can make it easier to prompt characters.

    empek17Jun 11, 2026· 5 reactions
    CivitAI

    First thing i noticed is that Anima turbo lora breaks on the v3 which kinda sucks.
    Otherwise model looks really good, it's nice to have models with updated character knowledge

    motimalu
    Author
    Jun 12, 2026· 1 reaction

    Thanks, yes the Turbo LoRA are little weaker but maybe still usable with a CFG of 2:
    https://civitai.red/posts/29155373

    luckmiriwindsJun 12, 2026
    CivitAI

    Model looks so good but I can't get good hands, any tips?

    motimalu
    Author
    Jun 13, 2026· 1 reaction

    Hey @luckmiriwinds, some tips that might help:

    negative prompt:

    - use "artistic error, bad anatomy, bad hands"

    positive prompt:

    - use "masterpiece, best quality, very aesthetic", preferably at the start of the prompt

    - use tags defining what the character is doing with their hands, e.g. a gesture like "double v", see: https://danbooru.donmai.us/wiki_pages/tag_group%3Agestures


    The model can be pretty creative and errors seem to mostly occur in parts of the generation that aren't described in a way that it understands.


    Using natural language after the tags string should also help to guide the prompt understanding because both the base and this finetune use a combination of tags and NL captions.

    I don't do this enough either in the previews, mostly only when there is something that can't be described with just tags - but just using an LLM to simply expand your tags into NL helps.

    motimalu
    Author
    Jun 13, 2026· 1 reaction

    Some testing of those suggestions in a post, can kind of see how the model might melt digits together or hide the hands entirely without positive/negative prompts to define how they should appear:
    https://civitai.red/posts/29170310

    luckmiriwindsJun 13, 2026· 1 reaction

    @motimalu Yea I realised hands are a lot about the negative prompt too! Thanks for the answer, love the model

    BoredafkJun 13, 2026· 2 reactions
    CivitAI

    I recommend also Expanding the Blue archive Data Cutoff(And maybe some others if there's any) It still Doesn't know Fuyu, Maybe Miyo(Haven't tested her that much), Rena, Ritsu, Even Kei

    Just a Suggestion

    BoredafkJun 13, 2026

    Oh. and the Umamusume one as well, If you can :)

    motimalu
    Author
    Jun 14, 2026· 3 reactions

    Hi, sure I can try to include these suggestions when/if I do a full finetune next

    BoredafkJun 14, 2026· 1 reaction

    Looking forward to it! Thanks for the Bangers!

    mac2492Jun 13, 2026· 2 reactions
    CivitAI

    It's here! Can't wait to play around with this when I'm back from vacation. Thanks as always!

    zjc1772665101270Jun 13, 2026
    CivitAI

    为什么这个模型下载量不那么多?

    motimalu
    Author
    Jun 14, 2026· 3 reactions

    ╮(ᵕ—ᴗ—)╭

    Kuro_ShurikoJun 13, 2026· 5 reactions
    CivitAI

    Will this still be updated if there's more base anima updates or will this be final?

    motimalu
    Author
    Jun 14, 2026

    I think so, if Anima has a significant update then yes it would probably be worth trying to finetune it too.

    FuturabbitJun 15, 2026· 5 reactions
    CivitAI

    I feel pity about myself finding this great model too late!

    goeypants400Jun 15, 2026· 6 reactions
    CivitAI

    finally someone who shows information of the training, really good detailed description

    shis63630252Jun 19, 2026· 11 reactions
    CivitAI

    我很喜欢你的模型，他的数据很新，绝区零的数据甚至可以截止到希希芙，并且直出也非常好看。

    uraotoJun 20, 2026· 10 reactions
    CivitAI

    The new model with additional character training is fantastic! Characters like Cure Arcana Shadow are rendered very clearly and accurately.

    However, I feel that the training for characters from Pokémon Legends: Z-A is not very good. Considering when the game was announced and the amount of available material, I would have expected the model to recognize them by now. Were those characters intentionally excluded from the training data?

    motimalu
    Author
    Jun 20, 2026· 3 reactions

    Hello, and thank you!

    The dataset is manually curated and many characters that might have readily available material will not necessarily be included as a result.

    So it might be more correct to say some characters not being included is a more of a result of them not being intentionally sought out rather than being intentionally excluded.

    uraotoJun 21, 2026

    @motimalu 
    Thank you for taking the time to explain it so thoroughly. It's a bit disappointing that those characters haven't been trained, but I'm glad to learn that this model isn't built through an automated process and that the data is carefully curated by hand.

    I can hardly imagine how much effort goes into that. I'm looking forward to seeing how the Kirazuri model continues to evolve!

    motimalu
    Author
    Jun 21, 2026· 2 reactions

    @uraoto No problem, and thank you for your understanding - in the future I will try to include these characters. ദ്ദി˶ｰ̀֊ｰ́)✧

    Fei_mauJun 22, 2026· 8 reactions
    CivitAI

    可惜我的顯卡不夠力順跑Anima

    老哥，你做得很厲害啊

    kdygbw2Jun 23, 2026· 2 reactions

    anima配置要求应该比光辉低一些吧

    king6sssJul 11, 2026

    假如你你的显卡有能力跑光辉，就有能力去跑ANIMA。你可以尝试使用INT8量化方法去使用ANIMA，它的显存要求更小，质量损失肉眼不可见。至于加速可以搜索torch compile编译或者直接使用蒸馏模型。

    necrophagism777Jun 23, 2026· 7 reactions
    CivitAI

    I think this is the best Anima fine-tune I have tried. Very new dataset, good style, and doesn't affect the natural language comprehension too much.

    lost1mongrel890Jun 27, 2026· 7 reactions
    CivitAI

    Hi there, will there be an int8 version in the future?

    motimalu
    Author
    Jun 27, 2026· 3 reactions

    Hello, thanks for the suggestion!

    I tried quantizing to int8 using convrot with https://github.com/silveroxides/convert_to_quant

    Looks like a ~20% generation speedup, and 4.2GB -> 2.3GB filesize with minimal degradation.

    Comparison: https://civitai.red/posts/29447606

    Added to hugging face for the moment (Civitai does not have an int8 model variant option):
    https://huggingface.co/motimalu/kirazuri-anima/blob/main/split_files/diffusion_models/anima-kirazuri-v3-int8convrot.safetensors

    Note: Int8 requires the latest ComfyUI and Ampere or newer Nvidia GPU.

    lost1mongrel890Jun 27, 2026

    @motimalu Appreciate it! I wonder how well this would run in T4 gpu

    motimalu
    Author
    Jun 27, 2026

    @lost1mongrel890 No problem - If you can try it, I'd be curious to know too

    1795142525Jul 8, 2026
    CivitAI

    为什么这个模型使用风格lora效果很差，是需要额外提高Lora权重吗

    motimalu
    Author
    Jul 8, 2026

    （机器翻译）简而言之——是的，某些 LoRA 可能需要更高的强度才能在此处发挥作用。

    LoRA 在训练它们的“基础模型”上效果最佳。

    此外，根据特定 LoRA 的超参数，它们的效果可能会有所不同。

    由于该模型经过训练后与基础模型存在一定差异，LoRA 的整体效果有所降低。

    作为参考，使用 LoRA 对 SDXL 系列模型进行逐步微调时，例如 SDXL -> Illustrious -> Noobai 等，也会观察到类似的效果。

    1164912321Jul 14, 2026

    3.0对lora的效果非常好，你得看看你lora有没有冲突，再看一下lora权重自己微调一下，

    Kuro_ShurikoJul 9, 2026· 1 reaction
    CivitAI

    Hello, since turbo version released will you also be doing the same for Kirazuri? kinda excited for it

    empek17Jul 9, 2026

    would also love it if we'd get a turbo version.
    This version with turbo lora i found to have a lot of problems with overall anatomy and compositions.

    motimalu
    Author
    Jul 9, 2026· 6 reactions

    Hello, at present I don't have a good picture of how to approach Turbo LoRA/Model training for Anima, though I am interested in trying.
    I will be releasing a new version 4.0 of this checkpoint soon which I hope will address some of the issues observed with anatomy and compositions though.

    motimalu
    Author
    Jul 11, 2026· 6 reactions

    A Turbo LoRA distilled from this checkpoint is available here now:
    https://civitai.red/models/2769422/kirazuri-anima-turbo

    Checkpoint
    Anima
    by motimalu
    Download (Beta) View on CivitAI
    digital illustration
    video games
    base model
    anima
    manga
    anime
    aesthetic

    Details

    Downloads
    2,474
    Platform
    CivitAI
    Platform Status
    Available
    Created
    6/11/2026
    Updated
    7/27/2026
    Deleted
    -
    Trigger Words:
    masterpiece
    best quality
    very aesthetic

    Files

    kirazuriAnima_v30AnimaBase1.txt

    Size:
    494.30 KB
    SHA256:
    bc036321336581420b6085b64903c1d253e71ff2331d21b84e3207ae00700cf9

    Mirrors

    CivitAI (1 mirrors)

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (192 mirrors)
    CivitAI (126 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    kirazuriAnima_v30AnimaBase1.txt

    Size:
    644.74 KB
    SHA256:
    8950c198eb5edc8c66fd6d4d28b8fd66bff6a2cf46b50a23fc88673965b800b0

    Mirrors

    CivitAI (1 mirrors)

    kirazuriAnima_v30AnimaBase1.safetensors

    Size:
    3.89 GB
    SHA256:
    3679ed7dc092b0c85c1b4b8d36b1741149a8ebcb0be832f71aeb801fe1d50003

    Mirrors

    HuggingFace (4 mirrors)
    CivitAI (1 mirrors)

    kirazuriAnima_v30AnimaBase1_txt.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (90 mirrors)
    CivitAI (117 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (2 mirrors)

    Available On (2 platforms)

    Same model published on other platforms. May have additional downloads or version variants.

    SeaArt
    Kirazuri (Anima) - v3.0 [anima-base-1]
    PixAI
    Kirazuri (Anima) v3.0 [Anima-Base-1] - v3.0 [Anima-Base-1]