CivArchive
    Indigo Furry Mix Anima - v1.0
    NSFW
    Preview 137053822
    Preview 137053821
    Preview 137053834
    Preview 137053831
    Preview 137053840
    Preview 137053843
    Preview 137053844
    Preview 137053847
    Preview 137053851
    Preview 137053853
    Preview 137053857
    Preview 137053858
    Preview 137053865
    Preview 137053863
    Preview 137053871
    Preview 137053873
    Preview 137053826
    Preview 137053825
    Preview 137053830
    Preview 137053835

    This is an experimental Anima model based on Anima base v1.0, finetuned on a very small amount of e621 and bara furry data. (Big thanks to a friend of mine who helped me train it!)

    So, being an Anima model, the architecture is not the same as SDXL. It runs on English natural language (NL) + Danbooru-style tags, and there's NO e621 data in the base. Even with my little finetune on top, it's still not great at e621 tags and concepts, the main improvement is the aesthetic style.

    Features:

    - indigoFurrymix_anima_v1 keeps most of the natural language ability and the text rendering (you can still prompt it to write words in the image most time).

    - It also keeps most of the original Danbooru artist styles, plus I've mixed in some bara furry artist styles (no promises on accuracy or how well they actually work, though).

    - Compared to Anima base v1, this model seems to react more to e621 tags — but I'd still recommend sticking with Danbooru-style tags.

    - That said, the natural language understanding on this model is very decent. Always feel free to just "describe" any missing e621 concept in plain English NL — sometimes it works shockingly well.

    - It's pretty obedient overall. It'll respond to both positive and negative prompts.

    - Compare to Anima base v1, The aesthetics got a light tune-up, so it hits a "usable" level even without artist tags. (Still, use style prompts or artist tags — you'll get better results.)

    - Can generate very dark image, and the stylization is very strong. Tag weights can also go very high — SDXL caps out around 1.8, but this model takes 2.5, even 3.

    Downsides:

    - Limb stability can be a bit shaky sometimes, especially finger and toe counts. Since it's missing e621's precise vocabulary like 4 toes / 5 toes, the model is basically just guessing. It also loves to add pawpads where you didn't ask for them.

    - Artist tags seem to have a pretty big impact on overall image stability, and background stability can be pretty rough sometimes, images can come out looking really muddy.

    - Some e621 concepts are just plain missing — it can't generate certain genital anatomy (canine sheath, genital slit, etc). And use 'testicles' instead of 'balls' (bth 'knot' will become real rope knot with anima xddd).

    - It can't do a lot of characters, Legoshi being one example.

    - Sometimes using 'anthro' without 'furry' tag will pull a full "Detroit: Become Human" on you — the character just turns human. So I'd recommend just throwing 'furry' into your positive prompt everytime.

    - And sometimes the aesthetic quality drops so low it makes me question my entire life. It definitely wasn't this bad when I was merging, so it's probably something I just didn't catch in testing :(

    TL;DR: not having e621 data really does hit hard.

    How to use:

    - You'll need WebUI Forge Neo: https://github.com/Haoming02/sd-webui-forge-classic/tree/neo or ComfyUI (I'm using portable): https://docs.comfy.org/installation/comfyui_portable_windows. Some of the preview images were generated with Forge Neo and some with ComfyUI, so the embedded metadata format won't be consistent across all of them.

    - You need to download the VAE, the CLIP/text encoder, and the diffusion model itself — all three, no skipping.

    - Sampler: Euler a or ER SDE. Keep steps under 30 (20–50 works, but honestly I don't think going higher does anything better). CFG 3–6 (I used 4 the whole time while testing).

    - Anima is great at generating high-res images natively — up to 1536px straight out of the box. I'd recommend going for 1k, no hires fix needed.

    - For artist tags, use the @xxx format, not by xxx. And don't mix more than two artist tags — any more and you'll start getting messy compositions or straight-up weird artifacts.

    - It feels 'really' different from SDXL to work with. Give yourself a bit of time to adjust anima based model.

    - It uses both the generic quality words like best quality, very aesthetic and a Pony-style score system. Don't forget to throw those quality tags in.

    - Artist tags: https://thetacursed.github.io/Anima-Style-Explorer/.

    Description

    - indigoFurrymix_anima_v1 keeps most of the natural language ability and the text rendering (you can still prompt it to write words in the image most time).

    - It also keeps most of the original Danbooru artist styles, plus I've mixed in some bara furry artist styles (no promises on accuracy or how well they actually work, though).

    - Compared to Anima base v1, this model seems to react more to e621 tags — but I'd still recommend sticking with Danbooru-style tags.

    - That said, the natural language understanding on this model is very decent. Always feel free to just "describe" any missing e621 concept in plain English NL — sometimes it works shockingly well.

    - It's pretty obedient overall. It'll respond to both positive and negative prompts.

    - Compare to Anima base v1, The aesthetics got a light tune-up, so it hits a "usable" level even without artist tags. (Still, use style prompts or artist tags — you'll get better results.)

    - Can generate very dark image, and the stylization is very strong. Tag weights can also go very high — SDXL caps out around 1.8, but this model takes 2.5, even 3.

    FAQ

    Comments (39)

    ValadionJul 18, 2026· 1 reaction
    CivitAI

    uhmm did you forget to upload the Diffusion model or is it just the text encoder and vae ?

    indigowing
    Author
    Jul 18, 2026· 1 reaction

    Civitai decided to make my model file disappear, so I had to re-upload it. It should be fixed now.

    CuauhtemocI5MALJul 18, 2026
    CivitAI

    Apparently, you tried to wait to see if someone makes an Anima version of NoobAI, or at least an Anima based checkpoint model trained with e621 tags but you ran out of patience.

    indigowing
    Author
    Jul 18, 2026

    No, it's because a friend of mine finetuned the aesthetics using some bara furry data with e621 tags, and I thought it looked pretty good, so I decided to give it a try — and see just how much of an impact the missing e621 concepts would have. I could obviously wait longer, of course.

    CuauhtemocI5MALJul 18, 2026

    @indigowing Ok, Is that finetune published here?

    indigowing
    Author
    Jul 18, 2026

    @CuauhtemocI5MAL Nah, because the finetune mixed in a lot of furry artist data, and I'm worried I'd get my ass kicked if I made it public. I've got a record already: some big artists (like dacad) have DM'd me before on X telling me not to use their arts for training, and a few kemono bara furry artists are still in contact with me right now. I can't release the separate weights, but once it's baked into the model I can still pass it off as "someone else trained it, not me, not my friends" :(

    YgjiyfchjjgfgJul 18, 2026· 2 reactions
    CivitAI

    This is so good 😻

    teren0894Jul 20, 2026· 3 reactions
    CivitAI

    Hi. I am trying to test this new anima checkpoint... but I don't know why all my pics are just... black background... nothing. Can someone help me?

    indigowing
    Author
    Jul 20, 2026

    You may need to provide more information, such as the prompt and parameters used to generate the image, otherwise it's impossible to determine what the problem is.

    FadedCani386Jul 21, 2026· 1 reaction

    Try turning off the VAE and keep only the text encoder active – see if that gets your images back to normal. I had the exact same issue: generation looked fine, but the VAE decode stage gave me black images. As soon as I disabled the Qwen Image VAE and relied on the text encoder alone, everything rendered correctly.

    teren0894Jul 22, 2026

    @indigowing Here is a post of mine... https://civitai.red/posts/29925798

    teren0894Jul 22, 2026

    @FadedCani386 I have thisd issue... even I put whole JSON on the post... https://civitai.red/posts/29925798 

    indigowing
    Author
    Jul 22, 2026· 1 reaction

    @teren0894 It says 'The page you are looking for doesn't exist' :o

    teren0894Jul 22, 2026

    @indigowing My Apologize, I wrote an article with the JSON workflow, the black square is the result that drops me... i dont know what I am doing wrong... Thanks for reply and your time for answer. Here is the link with the JSON attached: https://civitai.red/articles/32894/i-cannot-create-any-anima-pic

    shoes22Jul 24, 2026

    @teren0894 I told you to put qwen3 0.6b in Load clip, not indigo

    zjc1772665101270Jul 21, 2026· 3 reactions
    CivitAI

    Hi! Would you be able to make a LoRA for this? Thank you so much!

    zjc1772665101270Jul 21, 2026

    I created the highest-quality LoRA I could based on your model and its base model (approximately 675 MiB). However, due to copyright and related concerns, I assume I’m not allowed to share it freely—is that correct?

    indigowing
    Author
    Jul 22, 2026

    @zjc1772665101270 You are free to share any LORAs, as long as it is not used for any commercial purposes.

    uchihayochimi966Jul 21, 2026
    CivitAI

    第一次使用anima模型
    不太懂工作流的流程，所以就直接用了大大的工作流
    可是每次都卡在Clip_skip出error

    # ComfyUI Error Report

    ## Error Details

    - Node ID: 2

    - Node Type: CLIPSetLastLayer

    - Exception Type: AttributeError

    - Exception Message: AttributeError: 'NoneType' object has no attribute 'clone'
    看了網上的解決方法但還是弄不好

    indigowing
    Author
    Jul 22, 2026· 1 reaction

    好像anima不能用clip skip吧，我的例图工作流好像也没有clip skip设置啊，你直接把我的例图拖到comfy里应该也不会出clip skip？正确连法应该是：

    UNET加载器

    └── Anima 模型 → K采样器 MODEL

    加载CLIP

    └── Anima 的 base_txt.safetensors → K采样器 正面和负面条件

    ├── CLIP Text Encode（正面）

    └── CLIP Text Encode（负面）

    加载VAE

    └── qwen_image_vae.safetensors → VAE 解码

    空Latent → K采样器Latent → VAE 解码

    然后，我的工作流是用最近版本的comfyui portable弄的，可能跟其他分支的comfy不兼容，有可能出bug，也有可能是你comfyui或者节点插件不够新，反正你把comfy和easy use插件都更新一下试试，不行就用portable或者换别人的工作流

    uchihayochimi966Jul 23, 2026

    @indigowing 感謝大大，因為我是直接拖圖到comfyui查看工作流的

    mvqsrtzJul 26, 2026· 1 reaction

    anima不用skip，你去comfyUI里面找anima官方参考工作流，把里面的模型替换掉就行

    BlackFXJul 21, 2026· 2 reactions
    CivitAI

    And so it is here! Can't wait to test this in details

    BlackFXJul 27, 2026

    After a lot of tests, I'm... divided about the results. I compared the outputs to AnimaBase using the same parameters and prompts for both. While the training toward bara furries can be felt, there isn't really a stark difference with AnimaBase in the quality of furry characters; which I'm sure can be blamed on the lack of e621 dataset in these early version.
    However my biggest gripe with these model so far is that the outputs feel too "smoothed over" compared to AnimaBase, by which I mean that for many style and artists tags that would usually have a more sketch like appearance or even just visible pen strokes are completely smoothed over to a "clean appearance", which I think is regrettable because the ability to closely match many of those styles is imo one of the most important strength of Anima. Because of the tendency of AnimaBase to sometimes go overboard with some artist tags/styles this can superficially seem to give better results, I think the overall output quality would be better improved by mitigating this from the prompt rather than muting them during finetuning.

    BlackFXJul 27, 2026

    My previous comment can seem really negative because I mostly highlighted what I thought were the bad points of the model, but it's still a good model overall. I just think it could be even better

    indigowing
    Author
    Jul 28, 2026

    @BlackFX Thank you for the detailed feedback! The smoothing is intentional — though I get why it reads as a downgrade.

    The idea was to raise the floor rather than the ceiling. AnimaBase without artist tags falls back to a pretty plain default, and I wanted something usable straight out of the box. Tradeoff is exactly what you're describing: pulling the default up compresses things at the top, and the sketchy/rough styles get flattened the hardest since they sit furthest from that "clean" center.

    Some of it's just taste too, tbh. A lot of the sketch-leaning artist tags in AnimaBase come out rougher than I want for the stuff I mostly generate, and for JP bara anime-leaning furry work I think the cleaner default is a net positive. But I think where the balance should sit is something I'm still chewing on.

    Your point about mitigating from the prompt instead of during finetuning is fair, and it's definitely the more flexible route. It just moves the work onto every user every time, which is what I was trying to dodge. Different tradeoff, not a wrong one.

    If you want something closer to base behavior right now — someone extracted the diff between this and AnimaBase as a LoRA. Running that on AnimaBase at reduced strength (0.5–0.7 ish) should land you somewhere in between: some of the aesthetic baseline, but more of the base model's stroke texture left intact. Might be worth a shot. Indigo Furry Mix Anima v1.0 — AdaptiveHQ Extracted LoRA - v1.0 | Anima LoRA | Civitai

    And don't worry about it coming off negative, it's useful feedback. Also worth saying: the lack of e621 data is the bigger bottleneck right now, and that's where most of my attention is going. Style balance is on my mind but that one's further down the list.

    BlackFXJul 28, 2026· 1 reaction

    @indigowing I figured that the smoothing was a consequence of getting better base results. And on that there is no doubt you succeeded, the "out of the box" results are much better than AnimaBase

    That, and as you said the simple matter of taste make senses that this would push the model toward the "smoother" aesthetic. Nonetheless, I hope a better balance can be found without having to sacrifice the base quality.
    When doing testing with artists tags on previous generations of models, you could usually tell immediately that some artist tags would never work, as soon as they had anything like visible brushstrokes or penstrokes, or even just a "simpler" style. And it was a true breath of fresh air to test AnimaBase and see that this time the model COULD do all of that. It would be a shame for this capability to completely vanish.

    Placing the burden on the user is a good point, I'm more of a "power user" I guess, I like to spend a lot of time adjusting prompts and styles to get it "just right" (I even have a custom app to sample artist tag influences for each version of the model), but it's fair that this is not the case for every user, or even the majority of them

    The idea of using the distilled Lora is intersting, I'll have to try that and see what the results are.

    And finally, yeah the lack of e621 data is undoubtedly the most pressing concern for a furry model and successfully integrating it will no doubt have a strong influence on the style balance anyway

    zjc1772665101270Jul 23, 2026· 9 reactions
    CivitAI

    I converted this model into a LoRA.： Indigo Furry Mix Anima v1.0 — AdaptiveHQ Extracted LoRA - v1.0 | Anima LoRA | Civitai

    and INT8 ConvRot : IndigoFurryMix Anima V10 — INT8 ConvRot HQ Speed - v1.0 | Anima Checkpoint | Civitai

    fog1Jul 23, 2026

    hey have tutorial how convert ANY model and lora to anima and ofther

    zjc1772665101270Jul 25, 2026

    Yes. For Anima and other compatible models, use the official INT8 ConvRot tool:
    https://github.com/Comfy-Org/comfy-model-tools

    First run quant_int8_convrot.py with --dry-run, then perform the actual quantization. LoRAs normally don’t need to be quantized—keep them separate and apply them to the matching base model in ComfyUI. If necessary, merge the LoRA into the original BF16/FP16 model first, then quantize the merged model. Not every architecture is guaranteed to work, so always test the result.@fog1 

    fog1Jul 25, 2026

    @zjc1772665101270 a have version for page?

    hikaru1986Jul 23, 2026· 1 reaction
    CivitAI

    Got it too work with Neural Pixel, v.0.8.2, using Vulkan and a MI50 Instinct 16GB, nice. So, you don't to deal with ROCm shenanigans with older GPU.

    BoyKissserJul 23, 2026· 1 reaction
    CivitAI

    Can you explain more about the e621 dataset that you finetuned the anima base with? How many images are used, are they mostly male focused, etc

    indigowing
    Author
    Jul 24, 2026· 2 reactions

    The finetune my friend has running right now used around 50k images. It's not pure e621 concept data and it's aesthetics-first, a finetune with a strong bara kemono (Japanese-style muscular anthro) vibe, and it's almost entirely solo male anthro images. Next we're going to try an e621 concept finetune.

    zjc1772665101270Jul 24, 2026

    @indigowing wonderful

    socksosockJul 25, 2026
    CivitAI

    There are already attempts to bake e621 tags and concepts into ANIMA, and apparently it helps quite a bit sometimes. Might be worth trying to merge that into the model?

    Also, as far as I know, LAX is already prepping the NOOB V2 project. The dataset will likely include e621 concepts, so I think it's worth digging deeper...

    indigowing
    Author
    Jul 26, 2026· 3 reactions

    First one, the 160k e6 Anima LoRA is definitely worth trying (e6 artists and characters are awesome), but right now it seems to have contaminated Anima's own concepts. You can never get an actual weapon sheath out of it when typing “sheath” or “sword sheath”, it just always gives you an anatomical concept (penis sheath) instead, and same deal with “balls“, and still it can't do “erection from genital slit” properly. We're working on combining danbooru and e621 tags and building a replacement index for it. I'd like to try training our own concepts first with my friend (and hopefully it will generate both the penis sheath and the sword sheath correctly at the same time).

    Second one, I took a look at that noob v2 project. Right now it looks like it's just a character concept replacer rather than a full model, but it's worth keeping an eye on going forward.

    PawbedlyClawJul 30, 2026· 2 reactions
    CivitAI

    It's perfect.

    woshitiancaiba409Aug 2, 2026· 2 reactions
    CivitAI

    这个模型很棒。

    Checkpoint
    Anima
    by indigowing
    Download (Beta) View on CivitAI
    anime
    bara
    base model
    kemono
    furry
    male

    Details

    Downloads
    1,048
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/18/2026
    Updated
    8/5/2026
    Deleted
    -

    Files

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (214 mirrors)
    CivitAI (149 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    indigoFurryMixAnima_v10.safetensors

    Size:
    3.89 GB
    SHA256:
    2f57f2b218da13a5880ad092b03f0f61aa2b8b179972c75846e805dfa58bb2b8

    Mirrors

    CivitAI (1 mirrors)

    indigoFurryMixAnima_v10_txt.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (99 mirrors)
    CivitAI (136 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
    Indigo Furry Mix Anima - v1.0