CivArchive
    WAI-ANIMA - v1.0(base 1.0)
    NSFW
    Preview 132075530
    Preview 132075529
    Preview 132075532
    Preview 132075531
    Preview 132075533
    Preview 132075538
    Preview 132075539
    Preview 132075542

    (2) wai-anima1.0 - v1 - Free AI CHECKPOINT Download | Tensor.Art | Tensor.Art

    Based on BASE 1.0. If you encounter any issues, please leave feedback, and I will try to fix or improve them in future versions.

    Steps: 20-30

    CFG scale: 4-5

    Sampler: Euler A Normal

    How to Use

    You need to use ComfyUI or Forge Neo

    you need to download qwen_3_06b_base.safetensors for text encoder,

    and qwen_image_vae.safetensors for VAE.

    Example images use 1024×1344.Hires upscale: 1.5（Please limit it to no more than 1.5×.）, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,Denoising strength: 0.35~0.5

    Positive Prompt

    masterpiece, best quality,score_7,

    Negative Prompt

    worst quality, low quality, score_1, score_2, score_3, artist name,blurry, jpeg artifacts, lowres,censor

    This is the first version of the model and is still in the exploration stage. If there are any issues, feel free to leave a comment and I will try to improve it later. You can refer to the official parameters for guidance.

    Anima [Official] - preview3-base | Anima Checkpoint | Civitai

    How to Use

    You need to use ComfyUI or Forge Neo

    you need to download qwen_3_06b_base.safetensors for text encoder,

    and qwen_image_vae.safetensors for VAE.

    BASE ANIMA preview3

    Steps: 20-30

    CFG scale: 4-5

    Sampler: Euler A Normal OR ER SDE BETA

    Positive Prompt

    masterpiece, best quality,score_9, score_8, score_7,

    Negative Prompt

    worst quality, low quality, score_1, score_2, score_3, artist name,blurry, jpeg artifacts, lowres,censor

    Description

    FAQ

    Comments (165)

    Silith19527Jun 4, 2026· 5 reactions
    CivitAI

    tested a few random samples and felt much cleaner than the base anima with a thicker and more retro vibe.

    Adherence is better but seems less flexible in tweaking to other styles. Not many anima loras out there to just smack on so probably a bit more work to get the desired style out.

    jenya1337Jun 4, 2026· 1 reaction

    If there is not enough loras, create it yourself.

    GPUPoorChadJun 4, 2026
    CivitAI

    A bit early in my timezone yay :D

    Maybe could specify timezone next time though

    misoraJun 4, 2026
    CivitAI

    Anima Base 1.0 model was revealed a little while ago, but is the model revealed today (June 4th) a different model ? Or is it the same model?

    kainnosgoth1988Jun 4, 2026

    wondering the same here

    SyntheticSinFutaJun 4, 2026· 3 reactions

    If only anyone could read the very first line of the description
    "Based on BASE 1.0"

    Cult_of_the_Black_BuffaloJun 4, 2026

    It's a finetune of Anima 1.0, meaning someone took it and further trained it using their own dataset.

    OTAKKxJun 4, 2026

    Yes, the previous one was based on P3, the new is is Base.

    Yeon_Albus_Jun 4, 2026· 1 reaction
    CivitAI

    十分甚至九分的好用，对风格lora污染极小。

    不过似乎有一点过拟合，比如miku身边容易跟一只吉祥物，但是我没认真测所以也不确定

    ramastf0rce767Jun 4, 2026

    nova的也会出现，吉祥物多为小动物，比如铃仙的提示词有可能会出现兔子

    GuChengAXJun 4, 2026

    以前wai的光辉系列也有，洛天依会带着罗小黑

    pialesJun 4, 2026· 6 reactions
    CivitAI

    noticed that I get less AI looks and better anime style by not using the ( score tags ) and sticking to using the illustrious positives/negatives tags

    RaininosiJun 4, 2026· 1 reaction

    if you mean score_1, score_5_up, then yes, this only work with Pony SDXL. Neither SDXL nor Illustrious nor NoobAI do support it.
    The common positive tags for a checkpoint - if the creator didn't mention it - are usually masterpiece, best quality, very aesthetic, absurdres on the positive, on the negative I use low quality, bad quality, low details. I avoid "worst quality" and generally oversimplifying words like "perfect" or detail(ed) in the positive because it's kind of subjective and more than ever gives you random results you may not want to have.

    badlyfeJun 4, 2026

    @Raininosi https://huggingface.co/circlestone-labs/Anima#quality-tags

    straysnowJun 4, 2026· 2 reactions
    CivitAI

    For me, the biggest problem with Anima is that it can no longer draw the “fat mons”. It's a pity. I wonder if future versions will add data on this aspect, or if I should wait for some kind soul to create a relevant LoRA.

    I hope the next version will add related support. Using fat mons with WAI Illustrious will produce a great effect.

    notmyproblem324Jun 4, 2026

    It probably does it better, the difference is anima tag weights, try (tag:2) to :3 instead of max "(tag:1.3)"

    straysnowJun 4, 2026

    @notmyproblem324 I also tried adjusting the weight, even to 3, but it still didn’t work. Or does Anima’s fat mons require a different tag? Or a specific phrasing?

    straysnowJun 4, 2026

    @notmyproblem324 Anima seems to interpret "fat_mons" as general obesity. Increasing the weight affects the body shape, but there’s no change in the genital area. I tried searching for other keywords, but they had no effect either. I don’t know what to use to achieve it.

    notmyproblem324Jun 5, 2026

    @straysnow Weird, (fat mons:2) to 3 works for me, though I already have fat in negative regardless of nudity since this model seems to make really curvy characters. I'm using Comfy and CFG 4. A lora would be nice anyway.

    straysnowJun 5, 2026

    @notmyproblem324 I am using Forge Neo. Using (fat mons:2) just makes the character fat, but doesn’t make the private area more fleshy. In IL, it generates normally. QWQ

    NightBattlesRoomJun 4, 2026
    CivitAI

    Thanks for your hardwork,amazing job man!

    Hugh_MungusJun 4, 2026· 2 reactions
    CivitAI

    How in-built artist style this ck have?

    OTAKKxJun 4, 2026
    CivitAI

    Yes, June 4 is here finally, thanks for the best Anima model out there!

    CipherVisageJun 4, 2026· 3 reactions
    CivitAI

    In my testing, the character consistency is excellent. When using the same set of prompts, the character features remain almost identical, whereas the original anima base model tends to produce noticeable variations.

    Additionally, when applying a LoRA trained on the anima base, I noticed that some of the character's anatomical structures appear slightly distorted. I suspect this compatibility issue arises because the LoRA was not trained directly on anima-WAI.

    zhuangkechengnew209Jun 4, 2026
    CivitAI

    請問chloroform提示詞在這模型完全失效嗎?

    ProtoniumJun 4, 2026

    what the fuck

    degurshaftJun 4, 2026· 6 reactions
    CivitAI

    If anyone runs into issues training lora on this Anima merge, try renaming the layers key prefixes from model.diffusion_model to net. That helped me

    realjuanJun 4, 2026

    This fix worked for me as well

    shao_showJun 4, 2026
    CivitAI

    其实感觉，anima对自然语言的理解还是有点问题，还是说我的用法有问题，有没有hxd讲讲

    LittleCoogeJun 4, 2026

    丫用的那个qwen才0.6B你指望个啥，等有人炼更适配的高质量qwen再说吧

    NTR_BLACKJun 4, 2026

    那個提示詞模型才0.6B,你就別指望理解能力有多高,現在我覺得湊合可以用，已經算很不錯了

    換更高B的模型再訓練也許會更好,但是那個VRAM和性能要求絕對不像現在那麼低, 換成個4b就要由現在1gb 提高到4gb了，這單純是提示詞部分，可能也沒多大改善... 給你上14b 量化也要20gb vram，那vram沒30gb 你別想用了

    takoyamaad250Jun 4, 2026· 2 reactions
    CivitAI

    korewa sugoi

    Erin_Jun 4, 2026· 1 reaction
    CivitAI

    不知道是不是DIT架构或者是自己太菜的问题，总感觉Anima不适合Ultimate SD Upscale这种分块做highres的流程，容易出现一些奇奇怪怪的重绘纹理，面部修复也是，不知道有没有佬能指点一下

    NTR_BLACKJun 4, 2026

    我自己經驗是完全不需要臉部修復(我以前SDXL 光輝一定要修，不然眼不行，但新模型直接就出最佳表現)如果你有需要出高解像度---直接出2048X2048他都不會崩，DIT架構對異構解像度沒有什麼限制

    我自己是出1280X1024，然後加上升級模型提高到2.5K，5090D 7秒就能完成一張圖
    (推銷這個升頻超快不會出現偽影，不會清除掉原有的紋理，也不會過度銳化https://openmodeldb.info/models/2x-sudo-shuffle-cugan-9-584-969)
    X2不夠你就再接一個相同接點再跑一次

    你可以看看我的LORA分享內的樣張，是完全沒有套任何修復，分块highres

    花點時間在CFG STEP 和採樣器的配合上吧，只要原圖出得夠好，隨便找個基本模型放大器就可以出到4K 8K以上。另外我推薦加上Anima Turbo LoRA，CFG 2+0.7權重下出圖只要12-15STEP就跟之前差不多，而且仍然可以套上其他LORA，沒套這個我要30STEP


    Erin_Jun 4, 2026

    @NTR_BLACK 感谢，我确实没怎么试过直出大图，主要是习惯了A111的highres fix和Ultimate SD Upscale了...我的需求是出那种接近A4纸的2480*3508这种大小，以前的工作流就是靠Ultimate SD Upscale做到接近2k，然后用upscale模型放大到目标大小，这样能尽量减少伪影并且有一些画面细节

    NTR_BLACKJun 4, 2026

    @Erin_ 直接出大圖再X2是最簡單,前提是你的顯卡VRAM是足夠
    https://civitai.com/images/132809165
    我隨便測試了一張(我這LORA眼還沒訓練好,所以差一點)，你可以參考一下效果和參數
    1752X1240+2X的放大器，生圖時間13.5秒 (5090D),VRAM需求為11.5GB

    EvilFearJun 4, 2026
    CivitAI

    Good and bad, good with character but bad with background. It's the only model so far with Anima that give me that problem with backgrounds. Often solid color, do not follow the prompt, or give me a bad backgrounds. Tested same prompt with 4 different models (Anima), all followed the instruction for the background. Only Wai-Anima did not.

    EvilFearJun 5, 2026

    to be fair, the base model have a serious problem with overexposed light, which doesn't help.

    DC_ftwJun 5, 2026

    I'm having almost the same issue. The background, which generated without any problems in previous builds, is now forced to be a plain white “simple background” in the current build. Ofcourse, the background prompt is definitely there, and it worked fine before.

    MXtremeJun 5, 2026· 1 reaction

    using "solid background" in the negative prompt solved this issue for me

    DC_ftwJun 6, 2026

    @MXtreme didnt work for me... still having completely white simple background or white letter box

    YourmomdJun 4, 2026· 1 reaction
    CivitAI

    Does anyone else have problems prompting with characters like Kamisato Ayaka. It has the outfit and hair right but not the eyes.

    It used to work with Wai-Illustrious-NSFW

    Domia22Jun 5, 2026· 2 reactions

    some of the characters have to be prompted in different ways compared to illustrious models. I found this list of Anima tags. I opened the .csv in VSCode and just did a ctrl+F to search for characters names to find the differences in how the model expects them to be referred to. https://github.com/BetaDoggo/danbooru-tag-list/releases/

    SuccuponJun 5, 2026· 1 reaction

    Just a thought, if you are running detailers you can lose some of those very specific details of a character. You could try adding the character prompt to your detailer.

    Outside of that, try referencing this csv
    https://huggingface.co/datasets/Laxhar/noob-wiki/blob/main/danbooru_character_webui.csv

    This contains tags that also describe the character. (so things like the eye color in this example)

    You can add those tags in along with your character tag and you might get better results

    YourmomdJun 5, 2026· 1 reaction

    @Domia22 Bro thanks its working now. Also an easier way is you can use an AutoCompleter located in settings in ComfyUi, so you don't have to Ctrl+F everytime in VsCode and just type in the positive prompt box.

    YourmomdJun 5, 2026

    @Succupon I never dealt with detailers before actually

    Domia22Jun 5, 2026

    @Yourmomd I had no idea that was a feature lmfao this is actually life changing haha thanks man

    DenkiSolosShippudenJun 5, 2026
    CivitAI

    Does the knowledge cutoff of anima vs. illustrious drastically affect images, in y'alls experience, for newer characters? Like, do they look the same as the rest of the model's generations if they're "only from the anime side" of the merge or nah?

    kuraigJun 5, 2026· 2 reactions
    CivitAI

    this anima thing better than illustrious model?

    Tozi_WhiteJun 5, 2026· 3 reactions

    Anima is much better, the only downside for now is that there is less LORA but that will change a bit.

    DarkEngine2024Jun 5, 2026· 4 reactions

    There's much more versatility and prompt adherence, but you have to tweak and iterate the balance of style weights/loras to get the result you want. Illustrious is less versatile but much quicker to get results you want in a style you like.

    AI_DKJun 8, 2026

    prompt understanding is magnitudes better, though you don't get tag strength that XL models/clip has, so you need to rely on LoRA, and if none, you can just edit with inpaint.

    LazyMofoJun 8, 2026

    every time a new model comes out, it another learning curve, and bunch more computing power and time wasted to learn. I wouldn't put my hopes up unless you're not satisfied with the model you're currently using. Besides, another one may come out in just a few weeks. I'd wait til it reaches the tenth version and top of the chart before learning this "so much better model" yadayadayada...

    necrophagism777Jun 8, 2026

    pros: better prompt adherence, easy multiple characters, better action/dynamic scenes, way less overfitting: better varieties.
    cons: harder to prompt(simple or conflicted prompts = bad image), less lora, harder to mix style, no good controlnet yet, upscaling and inpainting are still not very stable.

    Elaina996Jun 9, 2026

    @necrophagism777 精辟！我以为是我个人操作的问题，放大和修复时好时坏的，很搞人心态

    CharbelJun 5, 2026· 5 reactions
    CivitAI

    Finally!!!!! 🎉 I wasn't completely sold on the previous versions, but the base v1.0 is undoubtedly one of the best Anima models I've ever tried. 🌟 My new LoRAs are absolutely on fire with this model! 🔥 Thanks!!! 🙌

    mrj666Jun 5, 2026· 1 reaction
    CivitAI

    老哥们，歪佬的base和anima的本体base有什么区别吗？最近太忙了，没时间玩

    lucusoeuJun 14, 2026

    画风更丰富一点，但是原版的动作提示词理解能力更强一些。我用下来，同一套提示词，原版出图的动作比wai的要更符合我的预期。

    plan_trusterJun 5, 2026· 10 reactions
    CivitAI

    ai-generated:1.5 slop as always

    xiangkao321Jun 5, 2026
    CivitAI

    Very Nice.

    sujie199519dw172Jun 5, 2026
    CivitAI

    请问什么时候能在c站训练器使用

    SurveillanceJun 5, 2026
    CivitAI

    I beg of you for an update to wai shuffle noob.

    ProtoniumJun 6, 2026

    no

    andandeai0608Jun 6, 2026· 4 reactions
    CivitAI

    I have question.When try to use anima_wai as base model to train a lora, the terminal show "missing keys in checkpoint", when i change base model as anima_base_1.0,the error missing.Aside from style, what are the differences between the two?

    Lifi_HafJun 6, 2026· 3 reactions

    you shouldn't train on merged checkpoint, doing so only end up sabotaging your own work.

    ProtoniumJun 6, 2026

    jeez, at least wait before training on this

    NTR_BLACKJun 6, 2026· 2 reactions
    CivitAI

    新版本我如果使用juo sena 十王星南，經常只要有pantyhose 相關提示詞就會出現那兩條大腿帶，但是PW3是沒有這個問題，我用LORA也沒辦法覆蓋他，我覺得應該是微調時把她的運動YOGA 褲上的那兩條白色帶訓練進去

    LittleCoogeJun 7, 2026

    用alternate costume这个tag消除原本服装

    PPPParadoxJun 6, 2026
    CivitAI

    训练lora时提示RuntimeError: Missing keys in checkpoint: ['x_embedder.proj.1.weight', 't_embedder.1.linear_1.weight', ...]。有人能训练到吗

    AlesteriaNo1Jun 13, 2026

    dont use this model in training. actually it's a simple merge of Yume and base anima (maybe some loras, idk). The author can't solve this problem, the model is disabled in training.

    findersJun 6, 2026
    CivitAI

    Don't know why but using it with "straw hat" prompt will generating oversaturated pictures.

    Pumba12Jun 6, 2026
    CivitAI

    does someone know how to get a specific anime style for example "one piece elbaph arc" ?

    ThongorJun 9, 2026· 4 reactions

    This kind of style is almost impossible without lora. Mainly because Anima checkpoints are trained on a large dataset that doesnt have anime screencaps in them.

    For general anime style, you better try https://civitai.red/models/2675648/bas-better-anime-style-plus-or-anima-base-v10-or-model-or-checkpoint?modelVersionId=3004325.

    q1925615622918Jun 6, 2026
    CivitAI

    为什么我出不了全身图，都是半身图

    fastdancesJun 9, 2026

    full body view,

    q1925615622918Jun 11, 2026

    @fastdances 我之前也加了类似的tag，他还是容易出半身图，后面我发现可能是一开始没有描述背景，他总是人物撑满全图，后面加上背景就可以了。

    IcoNx937Jun 6, 2026
    CivitAI

    Heyho,

    The checkpoint works great so far and eliminates some of the downsides that the anima_base_10 model has. However, I'm still running into a strange issue: sometimes the model generates monochrome images.

    When I put monochrome and greyscale in the negative prompt, it tends to generate characters with a bluish tint instead.

    aaa1324abcJun 7, 2026
    CivitAI

    示例图工作流里没有text_encoders的节点怎么运行的

    honyoripJun 9, 2026

    Since the sample image was generated by Forge Neo, dragging it into ComfyUI will not create a workflow for Anima that includes nodes for text_encoders. Please look for a workflow for Anima in the ComfyUI templates and use it.

    zjunchoon969Jun 8, 2026· 4 reactions
    CivitAI

    “high” “wide” base1.0这个版本会把人物全身的长宽比破坏掉，最明显的是 阔腿裤 "wide legs trousers " 会让人物明显变宽或者说变的过于丰满。然后high 也会有变化。 pw3 这个问题会小不少

    necrophagism777Jun 8, 2026· 5 reactions
    CivitAI

    It's a good merge/finetune. The image style looks more "modern" compared to the original base: lines are cleaner and color more smooth. Definitely not as flexible on artist styles as base, but that is expected. Just hoping that the "modernization" trend don't go too extreme like Wai-illustrious in later versions.

    gisaJun 9, 2026
    CivitAI

    Can anyone give a synopsis on how Anima is different from other models?

    lucusoeuJun 10, 2026· 1 reaction

    文字理解能力大幅提升，抽卡次数显著减少，能一定程度上理解句子，光辉模型的时代要结束了

    ZaltarJun 10, 2026· 2 reactions

    As @lucusoeu said, all this. But it's slower tho. I came from 4,90 i/sec on a 1152 x 896 to 1,90 i/sec. Illustrious is still working pretty well, but if you want to do more precise things without using C-net but simply by prompting, Anima is honestly impressive.

    FrankainsteinJun 10, 2026· 3 reactions

    it's like an illustrious model but that actually take natural prompting and is not asking as much vram and ressource as switching to flux/z image/Qwen and other bigger image models while still having a big jump in prompt adherance, and lora are way smaller like half of the size for the same quality so it's easier to stock up or make huge lora with big ranks and stuff x) it's still lacking in the fury departement and there is obviously less lora since it's news, and for specific 3D artstyle illustrious can still do better x) but for V1.0 base that it's even competing with our latest illustrious models is what make it look appealing. If your going to try you can just keep using your danboru tags setup at first then try natural sentence in it x)

    and it's uncensored too like illustrious no weird anaotmy stuff going

    oh also what is pretty cool as a lora creator is that i can make a dataset for both illustrious and Anima and have both danborus tags and natural sentence and it work x) i can use the same exact dataset on both ! so it's really nice comparé to making lora from illustrious to Z image and other it's a bit annoying and ask a lot more of restructuaration of tagging !

    razdel19Jun 10, 2026· 3 reactions

    It's much easier to achieve complex compositions and scenes with just a few tags and simple natural language. It's also much easier to train LoRAs in my experience, the model learns a lot from very few examples.

    On top of that, it needs somewhat less VRAM than Illustrious or similar models, though it is a bit slower.


    The model is still relatively new, so the community hasn't pushed it yet the way they did with Illustrious, which gave rise to NoobAI and similar models, IP-Adapters, and better ControlNets (we already have canny and scribble for Anima).


    I also feel it's a much more artistic model, it can replicate almost any style you train into a LoRA nearly to perfection.

    semiramisxxvJun 12, 2026

    i had the same question before. like what's the difference between illustrious and anima like to me i don't see a significant difference until i tried it myself and like what most people said. it's more to prompt adherence. personally i don't see which is superior than the other.

    say i want to generate a character with their long hair flowing down on her back. illustrious often failed to deliver it. most of the time the hair will be over their shoulder or flowing down on their chest instead. on anima it's just really accurate. whatever prompt you put in anima it follows whatever it is. add a little bit detail instruction it'll increase the accuracy for it.

    but as for me i still stick with illustrious. since illustrious have tons of fan-lora and checkpoint and commonly used to generate anime so illustrious still taking the throne for the time being. for anima lora and checkpoint aren't as many like what illustrious have. but if you want to experiment and hoping for accurate result give anima a try.

    NatzoJun 10, 2026· 4 reactions
    CivitAI

    Ah man, do I need to redownload new versions of all models? I barely kept up moving to Illustrious...

    MaxiimusJun 14, 2026

    yeah lol, I wish that they could just have a converter or something but thats probably not how it works yet

    Swagz438Jun 11, 2026· 6 reactions
    CivitAI

    Illustrious LoRas work with this?

    SOSUAIJun 11, 2026

    nah

    StraitjacketJun 11, 2026· 2 reactions

    It's not even SDXL, it's a completely different architecture. Outside some styles it doesn't need loras nearly as badly anyway.

    name12138Jun 11, 2026· 1 reaction
    CivitAI

    This model is absoluly exordinary.
    This model have a great ability of following and understanding prompts. In my attempt, it could generating 4 diffrent character with diffrent action, color, express and age. Which was the privillege of model with double clip and dozens B of parameter count.


    Moreover, the model supports many styles and famous characters, precisly generates of multi characters, and it has capacity of generating text, english and maybe japanese (I can't recognize are the letters true or not).

    However, as a fresh model, in my opinion the only defect is the unsupports of the community, such as controlnet.

    But as a lightweight model, it truely beyonds my expectations.

    And thanks to the creator that brings this extensive and accurate model.


    erasedc39149Jun 12, 2026· 1 reaction

    Would love to see the image you made. I'm trying to understand how to use multiple character positioning and posing myself

    MikadoNagiJun 11, 2026
    CivitAI

    This model sometimes draws female pubic hair even without being instructed to do so. Is there any way to stop this?

    PlumBeerJun 12, 2026

    Added female pubic hair to Negativeprompt.
    Have you tried this? Make use of negative prompts.

    MikadoNagiJun 12, 2026

    Adding female pubic hair to negative prompts will not have any effect.

    Instructions such as "She doesn't have pubic hair" will also not have any effect.

    PlumBeerJun 12, 2026

    @MikadoNagi positiveprompt (shaved pussy)/ Negativeprompt ：Place it immediately after the quality tag.　(female pubic hair:1.4)
    If this doesn't work, please send me the image. I'll try it too.

    zenova000Jun 12, 2026· 23 reactions
    CivitAI

    Since this is a NSFW model I will say it here. Seedance 2.0 just roll out a NSFW finetune called 「Hotgen Video V5」 through Byteplus' Japan subsidiary "Sousaku AI". You will have to be over 20 years old to be qualified to use this model under Japan law.

    zenova000Jun 17, 2026

    https://civitai.red/models/2708551/anima-tile-and-repair-controlnet-lllite?modelVersionId=3042408

    Anima tile controlnet is finally here! There is nothing Anima cannot create from now on, what an era of art we are living in!

    VelutinxJun 12, 2026
    CivitAI

    Any specific reason why ConfyUI or Forge is required?

    AlexMeowlerJun 13, 2026· 2 reactions

    because anima is not widely supported besides those 2 UIs
    p.s. with forge, its forge Neo, the default old forge does not support anima as well

    AlexMeowlerJun 13, 2026· 1 reaction
    CivitAI

    In Forge, there is generation setting "Shift" which is defaulted to 3 for Anima preset. Does it need to stay at that value?

    fredakan3047164Jun 13, 2026· 6 reactions
    CivitAI

    How can I get consistent art styles with this? Each output looks different with every batch, unlike Illustrious models that seem more consistent

    Akisy5Jun 14, 2026

    Adding a style lora might solve your problem.

    Akisy5Jun 14, 2026· 1 reaction
    CivitAI

    GOAT.

    I think everyone should try this model.

    jayjaytw88231Jun 14, 2026
    CivitAI

    I cannot use the new version. When I try to run your workflows from the example images it says

    CLIPSetLastLayer:
    'NoneType' object has no attribute 'clone'


    When I remove this node and connect the checkpoint CLIP links directly into the CLIP Text Encode nodes, it says:

    CLIPTextEncode

    ERROR: clip input is invalid: None If the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.

    PPPParadoxJun 15, 2026

    Clip loader need to download qwen_3_06b_base.safetensors

    jayjaytw88231Jun 15, 2026

    I already have it - in my ComfyUI folders models/clip and models/text_encoder. It does not help. Shouldnt it be part of the checkpoint itself ? Because the Load Checkpoint node from Wai's example images does not even load qwen_3_06b_base.safetensors

    jiam898989Jun 30, 2026

    @jayjaytw88231 Has it been solved？bro

    FaulmleyJul 4, 2026

    @jayjaytw88231 You can't use Load Checkpoint (Well you can but you probably shouldn't), the model isn't a checkpoint (despite being marked as a Checkpoint Merge). The checkpoint loader will load the diffusion model, text encoder model and VAE packaged in one file, but this model is separate, you download the provided VAE to the vae folder, the text encoder to the text_encoders folder, and this model to the diffusion_models folder. Then you use Load VAE, Load CLIP and Load Diffusion Model nodes to load the three parts separately. The Load Checkpoint can be used to load the diffusion model, but at that point it just can cause more confusion, best to just load diffusion model directly.

    FoxyTheWiseFoxJun 16, 2026
    CivitAI

    Does anyone have any prompting tips w Anima?

    Lynx2025Jun 16, 2026· 7 reactions

    https://www.alibabacloud.com/blog/model-studio-wan-video-generation-prompts-recipe_602777

    UtsurutteJun 17, 2026· 4 reactions

    The base anima model has info about how to prompt in the description.

    MeridisJun 17, 2026· 2 reactions

    I heard 'deviant art' in the negatives improving things, i tried and results are pretty good.

    zenova000Jun 17, 2026
    CivitAI

    https://civitai.red/models/2708551/anima-tile-and-repair-controlnet-lllite?modelVersionId=3042408

    Anima tile controlnet is finally here! There is nothing Anima cannot create from now on, what an era of art we are living in!

    gothwomenenjoyerJun 18, 2026
    CivitAI

    amazing model, just don't make my mistake and use the wrong text encoder 😂

    MrGeRoIJun 19, 2026· 2 reactions
    CivitAI

    Like always i still getting drips on every picture by WAI. In other this is best Anima model rn

    coralhymnJun 20, 2026· 1 reaction
    CivitAI

    It seems that this version will frequently have a white background. Even if words related to the white background are added to the negative prompt words, a white background will still appear. I'm not sure if it's a problem with the model or with my prompt words. The prompt words I used were "Dandadan" and "AyaseMomo".

    ytmmiJun 21, 2026

    可能是bese模型的问题，我使用anima_bese也经常出现白色背景

    nungmaJun 23, 2026

    It's anima thing. better prompt background to be something. or it will make a white or blank background.

    3cchi33v33Jun 20, 2026· 1 reaction
    CivitAI

    cant get this to work in ComfyUi, keep getting - clip set last layer nonetype object has no attribute clone ERROR

    JenkinsServantJun 20, 2026

    This isn't a checkpoint. This is a diffusion model.

    Basic start:
    - Load Diffusion Model (this model)
    - Load CLIP (qwen_3_06b)
    - Load VAE (qwen_image_vae)

    Don't use a clip set last layer node AKA do not try to implement any clip skips.

    RoyalMaidBelfastJun 24, 2026· 1 reaction

    same trouble. fixed when i put model in [ComfyUI\models\diffusion_models], before this i put it in checkpoint and unet =(.

    JenkinsServantJun 24, 2026· 2 reactions

    @RoyalMaidBelfast you'll learn more as you play around but if the models are all separate, then it probably isn't an all in one checkpoint like you are probably used to with SDXL. Although, I think some people have bundled their Anima models so they can be used in checkpoint nodes but I have yet to see it myself

    NTR_BLACKJun 21, 2026
    CivitAI

    如果作者下一次訓練可以考慮提升一下背景的表現就更好了，我覺得還是有一點單調，也許可以考慮引入一些GAL GAME的背景圖，但最好也同時提供提示詞範例

    另外就是眼睛很容易出現淚珠的問題還是沒有改善

    fapmindJun 21, 2026
    CivitAI

    As always the king of style prompt adherence. I rarely use anything else.

    Lynx2025Jun 21, 2026· 5 reactions
    CivitAI

    @WAI0731 does anima understand a JSONs with bbox coordinates ? I did try denos ideogram director and area prompts leaked to each other

    Lynx2025Jun 21, 2026

    https://civitai.red/images/134348590

    kayai_artJun 23, 2026· 4 reactions
    CivitAI

    I looove this model!
    Just one thing: On nsfw gens it does sometimes do blue male genitals? I even putted that into the negative prompt, but still. But its not always just sometimes.

    114514A48Jun 26, 2026· 1 reaction

    <(ºOº)>

    KINGOALJun 24, 2026· 19 reactions
    CivitAI

    Compared to Anima Base, this finetuned model offers significantly more stable outputs and much better default aesthetics. When paired with Turbo LoRA, it's an absolute productivity beast, cranking out high-quality images in a fraction of the time.

    However, there is a catch—this only works well if the model actually recognizes your prompt. I’ve noticed a significant loss of artist tags compared to the base model. Nearly half of the artist styles that were highly distinct and recognizable in Anima Base seem no effect or unrecognizable here.

    ·
    To give some context, my test is based on Anima-Style-Explorer. Given the massive scale, I sampled about thousand distinct artist tags that visually contrast with the model's default style. Conservatively, nearly half of them seem unrecognizable in this version.

    By "unrecognizable", I mean that when you look at the output, it is practically indistinguishable from a generation without any artist tags, whereas in Anima Base, you can notice a distinct visual shift compared to the default style.

    As I am not a model trainer myself and haven't cross-examined a wide range of other fine-tunes yet, I can't definitively say whether this is an issue unique to Wai's dataset or a broader systemic challenge common across the entire Anima fine-tuning ecosystem. It's just a notable bottleneck I observed during my own experience.

    Of course, this is not to say the model is bad at all—for the tags that do work, I actually really love the results and the aesthetic weight this model brings.

    kunde2Jul 2, 2026

    yea Anima turbo really seems to loose a lot of knowledge

    KINGOALJul 2, 2026

    @kunde2 I compared both Wai-Anima (with and without Turbo) against Anima Base (with and without Turbo), and the results are consistent: Wai-Anima itself has lost a significant amount of knowledge, and this regression is independent of the Turbo LoRA.

    From what I’ve observed, the main issue with the Turbo LoRA is that it degrades prompt adherence for complex descriptions and shifts the actual stylistic output of artist tags. It doesn't mean the model fails to "recognize" the tag entirely; rather, it recognizes the keyword but renders a little different (in fact, usually better) style compared to the non-LoRA base.

    That being said, I haven't done a comprehensive cross-examination across a wider range of fine-tunes yet. I’m still not entirely sure whether this severe knowledge loss is isolated to Wai-Anima, or if it's currently a common systemic bottleneck in Anima fine-tuning.

    a23153038Jun 25, 2026· 1 reaction
    CivitAI

    有沒有好心人能給個Comfy Ui的工作流，我不管怎麼試都都是錯誤

    xiaobai_solonJun 27, 2026

    把例图直接拉到comfyui上,就能得到作者的工作流了,多试几款,找一个结构比较简单不需要安装过多插件的,先上手试试

    ddt101Jul 15, 2026

    @xiaobai_solon 範例圖的工作流根本來的差不多但我用會報錯，沒辦法模型放大只能latent放大，得把原先的加載器分成unet加載器、加載clip還有加載VAE

    xiaobai_solonJul 15, 2026

    @ddt101 把工作流json和后台报错日志全发给AI,让它帮你分析下是缺少节点/模型/依赖/版本升级/硬件限制等等哪个问题?通常是可以诊断出来的.实在不行从comfyui的自带模版中搜索一下anima的基础工作流,那些都是最基础的官方节点,正常来说应该不会报错的.

    ddt101Jul 15, 2026

    @xiaobai_solon 把unet加載器又換回Checkpoint加载器（简易）就可以了，但clip跟VAE還是得另外搞他專用的加載器

    ddt101Jul 15, 2026

    @xiaobai_solon AI那個回答是真看不懂

    ddt101Jul 15, 2026

    @xiaobai_solon 好像是我其中一個需要VAE的沒有接到新的VAE加載器跟用不用Checkpoint加载器（简易 好像沒什麼關係

    xiaobai_solonJul 15, 2026

    @ddt101 Checkpoint加载器能用就行,比unet加载器慢几微秒,不影响出图,前者会检索一下模型自带的clip,unet加载器不检索,就这点差别,能用就好,没必要就别升级或新装插件或升级comfyui,容易依赖崩溃,更麻烦.

    luckysanJun 28, 2026· 4 reactions
    CivitAI

    It's a fantastic, top-notch model!!!!

    perverticaJun 29, 2026· 1 reaction
    CivitAI

    Best prompt adherence!!!!

    judas2991Jun 29, 2026
    CivitAI

    It's just me or PW3 looks better than base 1.0 in some cases? (more vivid colors as example)

    MrGeRoIJul 2, 2026· 5 reactions
    CivitAI

    Why i should use anime screencap instead of anime screenshot to make this look like anime?

    Carlson_fushiguro_PJul 6, 2026· 1 reaction

    Screenshot is messy and sometimes include unimportant things, while screencap is clean and only focus on the anime frame

    pialesJul 2, 2026· 4 reactions
    CivitAI

    the model tends to always do black or white background by default sometimes even when describing interior the roof is black even when i put simple background to negatives barely fix anything.
    could someone tell me if there's a way around this?

    Walden66Jul 5, 2026

    Emphasize the background prompt, such as with (room:2.0). Unlike with SDXL, a strength value of around 2 to 5 is appropriate.

    luckysanJul 4, 2026· 1 reaction
    CivitAI

    I'm getting an error when trying to train LORA on wai-anima.

    Is it currently impossible to run LORA training on wai-anima?

    x1101Jul 5, 2026· 7 reactions

    you train with base anima model then use this model for generating

    illiyasvielJul 8, 2026· 3 reactions

    Possible on diffusion-pipe, trained one for myself yesterday. But you will need A LOT of troubleshooting

    nsfrk3236973Jul 13, 2026

    Yes, this is a known structural issue. The wai-anima checkpoint is optimized for ComfyUI, meaning its internal tensor keys have extra prefixes (like model.diffusion_model.) that standard training frameworks like DiffSynth cannot recognize, causing them to fail to load the model.

    If you really need to train directly on the wai-anima checkpoint instead of the official base model, there is a workaround: You can just ask an AI (like ChatGPT or Claude) to write a simple Python script using the safetensors library to remove those extra prefixes and rename the tensor keys back to the standard DiT format. Once the keys are renamed, your training framework will accept it!

    Carlson_fushiguro_PJul 6, 2026· 8 reactions
    CivitAI

    Guys, is this One is the next checkpoint after illustrious? I just comeback here since the last year and seems like this one is very popular right now ngl. But most lora creators I know is quit in the last year tho, but maybe it'll have a great potential idk

    hakfull95539Jul 7, 2026

    Yes, I believe this represents a major leap forward for the new generation of image generation tools. I had taken a break for a few months, but seeing the results artists were achieving piqued my interest and made me curious enough to try it out. I’ve been testing it for a few days now, and honestly, it is truly amazing compared to the competition. Its standout feature is a far superior understanding of prompts and clearly better handling of LoRAs; I see it as a massive step up from what came before—especially considering the potential for further improvements and updates down the line.

    Hal171Jul 10, 2026

    It seems anima lora is more “user friendly” and stick to the style

    milo01Jul 6, 2026· 19 reactions
    CivitAI

    what is the different with this in the anima base?

    YourmomdJul 7, 2026· 3 reactions
    CivitAI

    Is it me or is there less variation in this model

    GtazJul 8, 2026· 24 reactions
    CivitAI

    nochekaiser881 territory

    zakazakacot125Jul 16, 2026· 2 reactions
    CivitAI

    Есть косяки с следованеим промта

    JossylessJul 16, 2026· 4 reactions
    CivitAI

    Interested even free buzz and no more fees, I like this Model Resources WAI-Anima

    FoxyTheWiseFoxJul 17, 2026· 2 reactions
    CivitAI

    Quick question about score tags on this one, do I actually need them, or is the model curated enough that it doesn't matter? I've seen people go both ways on this. Just wondering if there's a specific case where it actually makes a difference, or if it's basically down to preference at this point. Thanks!

    InoSimJul 19, 2026· 1 reaction

    Scores are interesting if you want danbooru scores. It search for voted results. score_1 is for extremely bad shaped drawings of arts, where score_9 is almost the best ever made images in danbooru.

    If you want to search for something specific, less scores are better than high scores. In reverse, if you want pure high quality image without specific situations/poses higher scores are better.

    You can also use high scores for controlnet upscaling which helps increasing bad scores first passes.

    SurveillanceJul 18, 2026· 4 reactions
    CivitAI

    I've been waiting for an update to wai-shuffle-noob from you. It's great for furry content. I hope you make a furry-focused, or, at the very least, a furry capable one of these in the future. <3

    HeetoJul 22, 2026
    CivitAI

    Thank you for WAI-ANIMA!!

    Checkpoint
    Anima
    by WAI0731
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    40,243
    Platform
    CivitAI
    Platform Status
    Available
    Created
    6/4/2026
    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)

    waiANIMA_v10Base10.safetensors

    Size:
    3.90 GB
    SHA256:
    9d5a1e1393c2978d6a979fab38fb0dee00bc2a94e354196c9f3cf2f6f56d5fbf

    Mirrors

    HuggingFace (26 mirrors)
    CivitAI (1 mirrors)
    TensorArt (1 mirrors)
    TensorHub (1 mirrors)

    waiANIMA_v10Base10_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.

    Tungsten
    WAI-ANIMA - v1.0(base 1.0)