CivArchive
    Chosen-mix_anima - v1.0
    NSFW
    Preview 138995271
    Preview 138995273
    Preview 138995266
    Preview 138995256
    Preview 138995267
    Preview 138995268
    Preview 138995300
    Preview 138995297
    Preview 138995254
    Preview 138995275
    Preview 138995281
    Preview 138995327
    Preview 138995276
    Preview 138995291
    Preview 138995295
    Preview 138995294
    Preview 138995304
    Preview 138995323
    Preview 138995331
    Preview 138995349

    The English instructions are below.

    大家好，我是chosen鲷

    这次是新搞了个anima大模型，我先用miaomiao1.4和one obesission融合，然后又把融合的模型进行微调训练得来了这个大模型，未来一段时间应该会主要用这个模型。

    画风和脸部加入了我自己的喜好，使用起来很简单，即便什么画风tag也不加，直接出图的效果也很好，然后它可以支持多画师串的输入，很多anima模型只能输入一个画师串，放多了图片会崩坏，这个模型没这问题，也就是说你可以像以前使用Illustrious模型那样去进行多风格融合，这点我很喜欢。

    我推荐的质量词和风格：

    1.封面图的风格：dutch angle, magical, evocative, highres, extremely detailed 8k wallpaper, illustration, depth of field, great quality, masterpiece, best quality, amazing quality, very aesthetic, absurdres, backlighting, detail eyes,

    @rella, @konya karasue, @an_yasuri, @kieed, @ruoganzhao, @yolanda, year 2023 BREAK,

    （请注意不要使用加速lora，不然出不来这个效果）

    2.masterpiece, score 7, highres, absurdres, best quality, explicit, chosen style 3, high-contrast aesthetic, sfw, safe, anime coloring,

    @akizero1510, @ye_jji,111

    最推荐这两个，还支持很多风格，大家具体的内容可以看例图中的tag。

    请额外注意，Civitai网站识别的图片信息并不完全准确，尤其是关于step cfg,我常用的两种搭配是cfg1 step12（使用加速lora的情况下）cfg4 step50（不使用加速lora的情况下），我会额外上传一个原图图包供大家下载原图，大家可以直接拖进comfyui里面读取图片的数据。

    然后有些画风如果你使用了加速lora的话是出不来的，必须直接出，封面图就是个很好的例子，一旦你使用了加速lora，就会失去那种朦胧感。具体的图片是否使用了加速lora，请大家根据例图中的开头是否有anima-turbo-lora-v0.2这个tag去判断。

    说实话，这个大模型我并不十分满意，因为我理想中的大模型应该有以下三个特质：

    1.易于使用，不需要复杂的tag，也不需要添加大量的lora就能出质量很高的图，稳定不易崩坏

    2.要能支持多种风格融合，能涵盖多种画风，能够直出大量的角色，且在美学表现力方面要足够优秀

    3.有优秀的兼容性，跟各种lora都百搭

    这个模型能做到前面两点，但是它第三点就做的不够好，它跟一些lora兼容性不够好，尤其是如果你使用了加速lora，再使用其他lora，请务必使用eular+simple的组合出图，否则就容易图面崩坏。

    这一点也是我未来版本更新的方向。

    感谢大家的下载使用，如果喜欢的话欢迎关注我，我的bilibili pixiv ID均为chosen鲷，欢迎大家关注我的频道https://space.bilibili.com/3867754?spm_id_from=333.337.0.0 https://www.pixiv.net/users/102907778 谢谢。

    The following content is from machine translation

    Hello everyone, I am chosen Diao.

    This time I have created a new Anima large model. I first merged miaomiao1.4 and one obesission, and then fine‑tuned the merged model to obtain this large model. I will probably be using this model mainly for the foreseeable future.

    The art style and faces have been adjusted to my personal preferences. It is very simple to use; even without adding any style tags, the direct output looks great. Moreover, it supports multiple artist string inputs. Many Anima models only accept a single artist string — if you add too many, the image will break. This model does not have that problem. That means you can mix multiple styles just like you used to with the Illustrious model, which I really like.

    My recommended quality tags and styles:

    1. Cover image style:
      dutch angle, magical, evocative, highres, extremely detailed 8k wallpaper, illustration, depth of field, great quality, masterpiece, best quality, amazing quality, very aesthetic, absurdres, backlighting, detail eyes,
      @rella, @konya karasue, @an_yasuri, @kieed, @ruoganzhao, @yolanda, year 2023 BREAK,
      (Please note: do not use the acceleration LoRA, otherwise you won't get this effect.)

    2. masterpiece, score 7, highres, absurdres, best quality, explicit, chosen style 3, high‑contrast aesthetic, sfw, safe, anime coloring,
      @akizero1510, @ye_jji,

    These two are the most recommended. There are many other styles as well; you can check the tags in the example images for details.

    Please note in particular that the image information recognized by the Civitai website is not entirely accurate, especially regarding steps and CFG. The two combinations I commonly use are: CFG 1, step 12 (when using an acceleration LoRA) and CFG 4, step 50 (without using an acceleration LoRA). I will additionally upload a package of original images for you to download. You can directly drag them into ComfyUI to read the image data.

    Also, some styles cannot be achieved if you use an acceleration LoRA; they must be generated directly. The cover image is a good example — once you use an acceleration LoRA, you will lose that hazy, dreamy feel. To determine whether an example image used an acceleration LoRA, please check whether the tag "anima‑turbo‑lora‑v0.2" appears at the beginning of the prompt.

    To be honest, I am not entirely satisfied with this large model, because my ideal large model should have the following three qualities:

    1. Easy to use — no complicated tags needed, and no need to add a lot of LoRAs to produce high‑quality images; stable and not prone to breaking.

    2. Able to support multiple style fusions, cover a wide range of art styles, directly generate a large number of characters, and have excellent aesthetic appeal.

    3. Have good compatibility and work well with all kinds of LoRAs.

    This model achieves the first two points, but it does not do well on the third. Its compatibility with some LoRAs is not good enough. In particular, if you use an acceleration LoRA together with other LoRAs, please be sure to use the Euler + Simple sampler combination to generate images, otherwise the image may break.

    This is also the direction for my future version updates.

    Thank you for downloading and using it. If you like it, feel free to follow me. My Bilibili and Pixiv IDs are both "chosen Diao". You are welcome to follow my channels:
    https://space.bilibili.com/3867754?spm_id_from=333.337.0.0
    https://www.pixiv.net/users/102907778
    Thank you.

    Description

    FAQ

    Comments (17)

    wohaishiyumiAug 7, 2026
    CivitAI

    ill版本还会更新吗老师

    chosen
    Author
    Aug 7, 2026

    那应该不会了，我现在都不咋用光辉模型了。所以4.1估计就是最终版了

    wohaishiyumiAug 8, 2026

    @chosen 太遗憾了。chosenmix是我用过最美学的模型

    chosen
    Author
    Aug 8, 2026

    @wohaishiyumi 你可以用anima模型了啊

    wohaishiyumiAug 9, 2026

    @chosen anima试了一阵子，但是总感觉相比ill，没那么顺手。老师你有工作流吗，我想学习一下

    chosen
    Author
    Aug 9, 2026

    @wohaishiyumi 这个模型下载页面里面有个压缩包，里面是原图，你只要拖进confy里就能读取工作流。其实我工作流没什么复杂的，就是简单地出图流程，连修脸都没挂。我刚开始用anima的时候也觉得不顺手，但其实是因为tag的使用方式不一样，原来的经验不适用了，但是你可以试试我这个模型，他可以像光辉一样进行多画师串融合

    wohaishiyumiAug 10, 2026

    @chosen 谢谢老师

    wohaishiyumiAug 10, 2026

    @chosen 老师，这个原图是jpg格式。导入comfyui中没有工作流欸。。。

    chosen
    Author
    Aug 10, 2026

    @wohaishiyumi 不会啊，我用插件把信息保存在图里了，我自己拖进去就能读取

    wohaishiyumiAug 10, 2026

    @chosen 真的没有欸。老师你有dc或者其他方式方便发工作流嘛

    chosen
    Author
    Aug 10, 2026

    @wohaishiyumi 我在可选下载页面添加了工作流的压缩包

    wohaishiyumiAug 11, 2026

    @chosen 太感谢了老师，实在是麻烦您了

    1q2w3e4rQAZAug 7, 2026
    CivitAI

    好耶

    movedcsxAug 8, 2026
    CivitAI

    这个模型综合来说让我印象深刻。画风很不错，但最特别的是它对于文本的还原忠实度来说比我用的其他微调模型感觉都更好（因为个人偏好原因我没有用太久基础模型，不知道基础模型的还原文本能力）。

    chosen
    Author
    Aug 8, 2026

    那个base模型，用自然语言效果会好很多，就是我经常看到一些老外发的例图，里面都是英语的大段句子，效果就很好，这点和光辉模型很不一样。我这个大模型，说白了就是把以前光辉那一套又给搬过来了，所以会感觉文本忠实度比较高。不过也是因为它忠实度太高了，导致只要tag固定，不管种子怎么变，生成的图像基本没区别，这个我有点不喜欢

    movedcsxAug 8, 2026

    @chosen 我已经是在用自然语言了（靠google或者本地小模型帮我翻译），不过也按原版模型里的推荐部分使用tag。奇怪的是，有些模型就是能不理某些自然语言或不理某些tag，最夸张时有些模型我能使用11强度的权重（就是在语言或者tag后面:11），而且有些模型还更容易生成土豆身体（完全没有手臂）。相比之下你的模型的重抽卡率让我觉得挺厉害的……相同的语言,几乎不需要重抽卡。我觉得anima模型的动作扩展性和手比光辉好了不少，但手臂在不少时候会直接消失……

    chosen
    Author
    Aug 8, 2026

    @movedcsx 我怀疑是训练集不够大的关系，所以导致只要tag固定，出来的图基本就一样，用了加速lora之后更加明显

    Checkpoint
    Anima
    by chosen
    Download (Beta) View on CivitAI
    anima
    illustrious
    comic
    anime
    base model

    Details

    Downloads
    622
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/7/2026
    Updated
    8/13/2026
    Deleted
    -

    Files

    chosenMixAnima_v10.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (220 mirrors)
    CivitAI (166 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    chosenMixAnima_v10.safetensors

    Size:
    1.11 GB
    SHA256:
    cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba

    Mirrors

    HuggingFace (103 mirrors)
    CivitAI (151 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (2 mirrors)

    chosenMixAnima_v10.safetensors

    Size:
    3.90 GB
    SHA256:
    04045f043f54a85493a499c087326561891c2f55c60b65de7253465449cb7a83

    Mirrors

    CivitAI (1 mirrors)

    chosenMixAnima_v10_trainingData.zip

    Size:
    30.68 MB
    SHA256:
    ca78dc834f443a0cc26e1d52921a6b924cfd63629c7a0eea6286438d05518ae9

    Mirrors

    CivitAI (1 mirrors)

    chosenMixAnima_v10.zip

    Size:
    13.39 KB
    SHA256:
    7c023f095a17d962097afbb01c4023fdd6bc4d7eb7f32cf26993482768a3dab1

    Mirrors

    CivitAI (1 mirrors)