CivArchive
    ✨ JANIMA ✨ - v1.0
    NSFW
    Preview 131507754
    Preview 131507070
    Preview 131505954
    Preview 131491262
    Preview 131500957


    👉 Join me on our discord server (https://discord.gg/tWrKk3jcXx)! Share art, ask for help, or just chat. We have CivChan bot plus games there!

    👉 Support future updates: Buy me a coffee or tip buzz! Every bit helps. Thank you for your support!


    JANIMA is a high‑quality text‑to‑image model finetuned from anima‑base‑v1.0, a 2B‑parameter anime‑focused diffusion model created by CircleStone Labs and Comfy Org.
    This checkpoint aims to keep Anima’s flexibility for anime and stylized illustration while pushing sharper details, cleaner anatomy, and more coherent compositions out of the box.

    JANIMA is designed for:

    • High‑detail anime characters and illustrations

    • Clean linework with crisp eyes and facial features

    • Richer backgrounds and small object details without heavy post‑processing

    • General anime/stylized art (not photorealistic/realism work)

    Model behavior

    • Inherits Anima’s strengths for anime and other non‑photorealistic content; realism is still not its main target.

    • Tends to produce more refined edges, textures, and micro‑details at the same resolution and steps compared to the base.

    • Works well as a “base” to stack LoRAs and style adapters on top, since it does not force a single over‑fitted style.

    Installing

    • JANIMA_v1.safetensors goes in ComfyUI/models/diffusion_models

    • qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders

    • qwen_image_vae.safetensors goes in ComfyUI/models/vae

    These are good starting points; adjust per your workflow and hardware:

    • Sampler: er_sde or similar modern sampler optimized for Anima‑based checkpoints

    • CFG: around 5 (you can push slightly higher or lower depending on prompt strength)

    • Steps: 24–30 (add more steps up to 50 if you are not planning to hi-res and upscale)

    Resolution: 512 - 1536 ^ 2 resolution supported.

    Quality tags (optional)

    You can prepend your prompts with quality tags like:

    masterpiece, highres, absurdres, newest, best quality, score_7

    These help bias the model toward cleaner, higher‑effort outputs.

    Suggested negative prompt

    A solid general‑purpose negative prompt for JANIMA:

    worst quality, low quality, lowres, score_1, score_2, score_3, blurry, jpeg artifacts, , long fingers, sepia, bad anatomy, missing fingers, watermark, artist name

    Tweak as needed.

    Helpful Resources

    To get the most out of JANIMA and the Anima framework, the following community resources are highly recommended:

    • Recommended ComfyUI Workflow: Anima Preview Workflow on Civitai

    • Style Reference: Anima Style Explorer – A visual database of over 40,000 artist tags and weights recognized by the Anima model.

    • Character & Tag Guide: Animadex – A comprehensive list of characters and tags with visual examples optimized for Anima generation.

    Notes

    • Best suited for anime and stylized illustration workflows (ComfyUI or similar) using the Anima ecosystem.

    • All preview images are raw model outputs (no over‑editing), so you can see what JANIMA actually does at reasonable settings.

    Description

    FAQ

    Comments (41)

    janxd
    Author
    May 26, 2026· 48 reactions
    CivitAI

    JANIMA v1.0 is out!

    You can join our discord to get v1.0 for FREE! If you are already in our discord, check the #updates channel for the links.

    If you wish to support me, you can buy me a coffee here: https://ko-fi.com/janxd It helps fund my trainings and run our awesome community!

    Workflow I recommend:
    https://civitai.com/models/2426853/anima-preview-workflow
    Other useful resources:
    Anima style explorer: https://anima.mooshieblob.com/
    Anima character and tag list with examples: https://animadex.net/

    Thanks for all the support! Looking forward to see what everybody makes ❤️🚀

    NiderJun 11, 2026· 10 reactions
    CivitAI

    My jaw dropped irl when I tried this model yesterday. It's insane how good it is at drawing multiple distinct characters on one image out of the box, without any masks or controlnets or other fancy toys. And don't get me started on intersecting tags. Your OC holding both a sword and a pistol, with another pistol in the belt? Good luck pulling allat off in pure JANKU, but this model easily handles such tasks with a proper tags+prose prompt.

    GombaTheGreatJun 11, 2026· 11 reactions
    CivitAI

    JANIMA is without a doubt currently one of the best anima models for image2image compositions. It picks up color and linework well from base images while keeping the character consistent. Artisttags work well too.
    I appreciate all the time and effort that went into making the model perform so well.

    Thanks @janxd ! 💖

    ClownReplicaJun 12, 2026
    CivitAI

    What settings and upscaler are recommended for Hi-Res Fix on Anima?

    EDIT: I'm assuming Hi-Res Fix is use, right? Please correct me if I'm wrong.

    prozombieJun 13, 2026· 1 reaction

    you don't need HR fix for Anima. it's not an SD model.

    if you need the image to be higher res, just generate at a higher res. Anima can handle up to 1536²px.

    ClownReplicaJun 13, 2026

    @prozombie I'm just asking because some of the example, like the March 7th one, are at far higher resolutions (1664x2432), but the metadata states it was generated at 832x1216. I assume this was achieved through hi-res fix.

    prozombieJun 13, 2026

    @ClownReplica it's probably just ESRGAN. here is an image I made with the same settings and everything, and upscaled 2x with ESRGAN: https://civitai.red/images/133675544

    in terms of fidelity, the output is almost exactly the same.

    demonkingoftyrannyJun 15, 2026

    i recommend upscaling with illustrious, otherwise use tiling like USDU. Anima sadly does not support sampling on large resolutions, so you can not do a second pass on the full upscaled image with that. But Illustrious in an img2img pass has no maximum size you can sample on arbitrary large images with low denoise in a second pass, anima gives you a completely distorted and pixelated image when you do that.

    demonkingoftyrannyJun 15, 2026· 1 reaction

    @prozombie you do need it for anima and illustrious also supports 1536^2 which is small, you want to upscale it and do a second pass on the upscaled image.

    prozombieJun 16, 2026

    @demonkingoftyranny we are talking about one-shot t2i here, not i2i.

    demonkingoftyrannyJun 16, 2026· 1 reaction

    @prozombie that is not how you do upscaling or "hires.fix", you do the first pass on the empty latent, then you upscale with a upscale model, downscale it to your desired resolution and then you sample again on that upscaled image. That is the standard process for a "hires.fix", you can not do that in one pass, the 2nd pass is always a img2img pass. But since you have all in one workflow it is one time clicking "run" or if you use Forge, A1111, or whatever it is also one time clicking run, but it are 100% at least 2 passes while the 2nd pass is an img2img pass on the upscaled image, that is how it works. And as i said i recommend using illustrious for the upscaling pass, it is also much faster, otherwise you would have to use tiling, there are multiple options, but i recommend Ultimate SD Upscale, since when you only sample on a tile you do not hit animas limit, but anima is very slow, so upscaling the anima image with illustrious is better imo.

    prozombieJun 16, 2026

    @demonkingoftyranny cool cool. not the point.

    HR fix isn't just an upscaler. it's a FIX to make SD/XL models not look like shit. it's a required step for those models.

    and that step is not required for Anima.

    demonkingoftyrannyJun 16, 2026· 1 reaction

    @prozombie it is just a second sampling pass on the upscaled image nothing more, no special fix. If you use simple tools like forge they just call it that way and hide from you what they do. They upscale the image and do a img2img sampling pass on it, a 2nd pass. Unupscaled base generation always looks shit, no matter if anima or illustrious, anima base generation also does not look better than illustrious. they are equally blurry, pixelated and lowres, you need upscale with 2nd pass for both to look good. But it is required for none of them, but both need it to look good.

    ln88888222Jun 13, 2026· 11 reactions
    CivitAI

    1、优点：3P、4P 姿势特别多，比其他优秀、灵活、自由。
    2、非常容易出现上下黑边、边框。而且无法用负面提示词抑制。

    acn00269Jun 21, 2026

    黑框是从pony时代就有的问题，大概率是图片设置的长宽问题，最好这些数字能被64整除，比如640x1280，多余的像素会被分块模糊处理，形成单色边框比如黑框灰框

    ReximflyJun 28, 2026

    大佬，这个模型对各种不同画师风格的适配度怎么样，是不是还是要带Lora才好用

    ln88888222Jun 28, 2026

    @acn00269 1024、1280也是一样会出现，没办法

    ln88888222Jun 28, 2026· 1 reaction

    @Reximfly 我喜欢的wada、深崎暮人还是可以的，其他不太好说 

    samson4617Jun 24, 2026· 2 reactions
    CivitAI

    I found that somehow it has some problems when working with anima turbo and the Euler Sampler. Everything is fine when using the default ER SDE Sampler with steps 25 or 30.

    Not sure if it's my own problem.

    penguin235Jun 25, 2026· 4 reactions
    CivitAI

    aesthetically it's quite nice but man you lose SO much of the prompt adherence capabilities of Anima by using these checkpoints

    Soldier177Jun 27, 2026· 2 reactions
    CivitAI

    Awesome checkpoint! But I have a question: Which Clip Skip is best to use?

    springmushroom_86Jun 28, 2026· 3 reactions

    DiT models like Anima doesn't use Clip Skip anymore.

    Soldier177Jun 28, 2026

    @springmushroom_86 Thanks. Actually, experiments confirmed this; there is almost no difference, and sometimes artifacts were even introduced.

    MaruzenskyJul 8, 2026· 1 reaction

    Anima doesn't use Clip Skip. I noticed Anima models skips the first 2 steps when the generation starts.

    BesssJun 29, 2026· 2 reactions
    CivitAI

    Im on Comfyui and im unable to use this model like IL checkpoint models i usually use, is it mandatory to use external clip model?

    BesssJun 29, 2026· 1 reaction

    UPDATE : just learned a lot of stuff, and apparently it doesnt work exactly the same as SDXL. so...
    Yea already figured things out.

    dev4insta801Jul 1, 2026

    @Besss Can you share what you figured out? I'm kind of new to this and want to use this on Comfy

    BesssJul 1, 2026

    @dev4insta801 Alright
    ill explain the practicals first then the theory yaps afer

    How to use Anima Models (Comfyui) :
    1. Loading the Base :
    Use the "Load Diffusion Model" Node.
    also, the model should be placed at "ComfyUI\models\diffusion_models".
    2. Loading Clip :
    Use the "Load Clip" Node.
    The model should be "qwen_3_06b_base.safetensors" [download is available at this page(Required Components)]
    this goes to "ComfyUI\models\clip"
    3. Load VAE :
    Use the "Load VAE" Node.
    The VAE will be "qwen_image_vae.safetensors" [download is available at this page(Required Components)]
    this goes to "ComfyUI\models\vae"

    The rest of the workflow works as usual,

    ---

    afaik, the difference between anima and other bases (e.g: SD, SDXL) is mainly about the prompt understanding. In SDXL, the prompts are understood best when it is written in what i'd call "Danbooru-style tags", which might feel kinda clunky time to time. Anima on the other hand, uses a LLM(Large Language Model] in this case qwen, As the adaptor of your prompt, this allows the model to understand a more descriptive prompts like "her multicolored hair streaked in pink and blue catching faint glimmers of neon light" And the model will still understand what it means.

    ***
    The downside : Qwen as the adaptor is kept frozen and general, after it had been taught the image side to fluently speak whatever the model outputs. While this keeps the model stable, it disallows growth, since it wont receive any more training(this might be little wrong, idk how to explain this part of the sentence accurately in english).

    ***

    pointless yaps :
    Because of that downside, some ppls tried to use 0.6B qwen to 4B one, but apparently it was a bunch of hassle and the output wasn't really "a big wow". So...
    whatever, JANIMA is 🔥 IG.

    dev4insta801Jul 1, 2026

    @Besss I have been doing this for a week or two now so I understand what you said. Would you be open to connecting on discord? I'm currently experimenting with Anima and it'd be good to connect with someone to discuss stuff like this. Regardless, thank you for the detailed answer :)

    BesssJul 1, 2026

    @dev4insta801 Sure, my id's "Bess18", im open to discuss at any time, just hit me up anytime you feel like it.

    dev4insta801Jul 2, 2026

    @Besss sent a discord request from "batmite"

    Soldier177Jul 1, 2026· 2 reactions
    CivitAI

    Friends, I apologize again, but I noticed that ANIMA models, including this one, do not respond well to token weights. What parameters do I need to add weight to so that the model can still use it? For example, I sometimes have to lift the weight to (-token-:4) for it to work.

    darkraven301Jul 2, 2026· 1 reaction

    This is how anima works here's a part of description from official one:

    Prompting

    The model is trained on Danbooru-style tags, natural language captions, and combinations of tags and captions.

    Use lowercase for tags, and spaces instead of underscores. Score tags are the only tags that use underscores.

    Recommended positive prefix: "masterpiece, best quality, score_7, safe, "

    Recommended negative: "worst quality, low quality, score_1, score_2, score_3, artist name"

    When using a tag that is different between Danbooru and Gelbooru, prefer the Gelbooru version.

    Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"

    Soldier177Jul 2, 2026

    @darkraven301 "Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"

    I understood that, but how high can you raise the weight of the token?

    richerichJul 9, 2026· 2 reactions

    @Soldier177 Normally, you can go up to :4 or even :5 in some cases without much issues (depending on how many tokens you already using, tho).

    kiheromasterki849Jul 2, 2026· 4 reactions
    CivitAI

    this model have good adherence, but it create extra hand sometimes

    LuckyRubberDuckyJul 5, 2026· 6 reactions
    CivitAI

    this model has excellent prompt adherence. good job.

    Sara_Jul 10, 2026
    CivitAI

    Best anima model for creating various styles, other models seem so force anime a lot, this doesn't that much. Would love to see you update this model or release the dataset for it's training.

    nellefrJul 13, 2026
    CivitAI

    At first i thought this model is pretty average. Turns out its leans towards tagged author style really well. I'll recommend using it with anima styles explorers services in mind.

    Hugh_MungusJul 14, 2026· 5 reactions
    CivitAI

    I love this but Anima is so hard sometimes...Barely respect my prompts....

    YUNMING98Jul 14, 2026
    CivitAI

    给到夯

    Checkpoint
    Anima
    by janxd
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    11,385
    Platform
    CivitAI
    Platform Status
    Available
    Created
    5/26/2026
    Updated
    7/26/2026
    Deleted
    -

    Files

    JANIMA_v10.safetensors

    Size:
    3.90 GB
    SHA256:
    38694ed21a93dd425fd3bf5a8c7a9f15070c1ecf6e5ab5e2138989c334ceea26

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (1 mirrors)

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (190 mirrors)
    CivitAI (125 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    JANIMA_v10.safetensors

    Size:
    3.90 GB
    SHA256:
    09c48368ae410664d2cc1d6d163edf79a75ca4101529739afe786dd5456f884d

    Mirrors

    CivitAI (1 mirrors)

    JANIMA_v10_txt.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (89 mirrors)
    CivitAI (116 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (2 mirrors)

    Available On (1 platform)

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

    SeaArt
    ✨ JANIMA ✨ - v1.0