最近试了一下 anima-preview2，体验挺特别。它最大的优势是能理解自然语言，也就是我们平时正常说话的方式。

以前用一些偏标签体系的模型时，提示词一般要写成这样：

Sousou no Frieren, long purple hair, purple eyes, black coat, white dress, puffy sleeves

这种写法对单人立绘还算好用。场景一复杂，问题就开始出现。

列如想描述一个第一人称视角的公主抱场景，角色陷入主视角的怀抱里。许多模型会忽略 pov，直接多画出一个莫名其妙的人去抱角色。再列如 sink in，模型可能只抓到 sink，开始画水槽、洗手池之类的东西。

anima 对这类自然语言的理解会好许多。它既能吃 tag，也能吃普通句子，出图更接近你真正想表达的画面，有时甚至会比预期更顺。

我这里使用的是秋叶大佬的 ComfyUI 整合包，安装 ComfyUI 的部分就不展开了，下面直接讲模型部署和基础工作流。

下载模型文件

打开 Hugging Face 页面：

https://huggingface.co/circlestone-labs/Anima

进入页面后，点击 Files and versions。

下面那些文件先不用管，找到 split_files 文件夹。

进入后，下载里面的 safetensors 文件。

下载完成后，把文件放到 ComfyUI 对应的模型目录里。

如果忘了 ComfyUI 文件夹在哪，秋叶启动器里可以一键跳转，直接从启动器打开对应目录会省事许多。

放好文件后，启动 ComfyUI。左侧模型列表里就能看到刚才下载的几个文件。

Anima 和普通 Checkpoint 的区别

把 anima-preview2 拖出来后，会发现它用的是 UNETLoader，不是常见的 Checkpoints。

这里需要注意一下。

普通 checkpoints往往把模型本体、VAE、CLIP 都打包在一起。加载checkpoint后，VAE 和 CLIP 一般会自动带出来。

anima-preview2 不是这种结构。它需要手动加载：

anima-preview2.safetensors
qwen_3_06b_base.safetensors
qwen_image_vae.safetensors

其中：

anima-preview2.safetensors → 用 UNETLoader 加载

qwen_3_06b_base.safetensors → 用 CLIPLoader 加载

qwen_image_vae.safetensors → 用 VAELoader 加载

提示词输入端必须接入 Qwen 的 CLIP。

图片输出前必须接入 Qwen 的 VAE。

Latent 部分照常使用，用来控制图片尺寸。

基础工作流搭建

一个最基础的 Anima 工作流，大致由这些节点组成：

UNETLoader
CLIPLoader
CLIPTextEncode 正面
CLIPTextEncode 反面
EmptyLatentImage
KSampler
VAELoader
VAEDecode
SaveImage

连接关系大致是：

UNETLoader 接到 KSampler 的模型输入。
CLIPLoader 分别接到正面和反面提示词节点。
正面、反面提示词节点接入 KSampler。
EmptyLatentImage 接入 KSampler 的 Latent 输入。
KSampler 输出接到 VAEDecode。
VAELoader 接到 VAEDecode。
VAEDecode 接到 SaveImage。

尺寸依旧在 EmptyLatentImage 里设置。列如可以先用：

864 x 1024

采样参数可以先按作者推荐来，也可以用一个比较稳的基础值测试：

steps：30
cfg：4
sampler：euler_ancestral
scheduler：simple
denoise：1

先跑通最重大。能正常出图后，再慢慢调参数。

使用现成工作流

如果不想手动搭节点，可以直接使用现成的 .json 工作流。

把下列json复制新建保存后拖进下图所示的文件夹内

{"id":"0a878181-81cc-4789-a625-e80873a54e14","revision":0,"last_node_id":15,"last_link_id":18,"nodes":[{"id":6,"type":"VAEDecode","pos":[631.2110708612264,-322.0564783734946],"size":[140,46],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"Latent","name":"samples","type":"LATENT","link":3},{"localized_name":"vae","name":"vae","type":"VAE","link":4}],"outputs":[{"localized_name":"图像","name":"IMAGE","type":"IMAGE","links":[5]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"VAEDecode"},"widgets_values":[]},{"id":12,"type":"EmptyLatentImage","pos":[-225.54409377604063,-123.71526100459779],"size":[270,106],"flags":{},"order":0,"mode":0,"inputs":[{"localized_name":"宽度","name":"width","type":"INT","widget":{"name":"width"},"link":null},{"localized_name":"高度","name":"height","type":"INT","widget":{"name":"height"},"link":null},{"localized_name":"批量大小","name":"batch_size","type":"INT","widget":{"name":"batch_size"},"link":null}],"outputs":[{"localized_name":"Latent","name":"LATENT","type":"LATENT","links":[11]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"EmptyLatentImage"},"widgets_values":[864,1024,1]},{"id":8,"type":"SaveImage","pos":[1307.2868894014575,-392.2849929458475],"size":[270,270],"flags":{},"order":8,"mode":0,"inputs":[{"localized_name":"图片","name":"images","type":"IMAGE","link":5},{"localized_name":"文件名前缀","name":"filename_prefix","type":"STRING","widget":{"name":"filename_prefix"},"link":null}],"outputs":[],"properties":{"cnr_id":"comfy-core","ver":"0.11.1"},"widgets_values":["ComfyUI"]},{"id":2,"type":"UNETLoader","pos":[-224.4966650980308,-256.749126772841],"size":[270,82],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"UNet名称","name":"unet_name","type":"COMBO","widget":{"name":"unet_name"},"link":null},{"localized_name":"数据类型","name":"weight_dtype","type":"COMBO","widget":{"name":"weight_dtype"},"link":null}],"outputs":[{"localized_name":"模型","name":"MODEL","type":"MODEL","links":[17]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"UNETLoader"},"widgets_values":["anima-preview2.safetensors","default"]},{"id":4,"type":"KSampler","pos":[162.38611852111273,-291.8327429355817],"size":[270,474],"flags":{},"order":6,"mode":0,"inputs":[{"localized_name":"模型","name":"model","type":"MODEL","link":17},{"localized_name":"正面条件","name":"positive","type":"CONDITIONING","link":8},{"localized_name":"负面条件","name":"negative","type":"CONDITIONING","link":9},{"localized_name":"Latent图像","name":"latent_image","type":"LATENT","link":11},{"localized_name":"种子","name":"seed","type":"INT","widget":{"name":"seed"},"link":null},{"localized_name":"步数","name":"steps","type":"INT","widget":{"name":"steps"},"link":null},{"localized_name":"cfg","name":"cfg","type":"FLOAT","widget":{"name":"cfg"},"link":null},{"localized_name":"采样器名称","name":"sampler_name","type":"COMBO","widget":{"name":"sampler_name"},"link":null},{"localized_name":"调度器","name":"scheduler","type":"COMBO","widget":{"name":"scheduler"},"link":null},{"localized_name":"降噪","name":"denoise","type":"FLOAT","widget":{"name":"denoise"},"link":null}],"outputs":[{"localized_name":"Latent","name":"LATENT","type":"LATENT","links":[3]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"KSampler"},"widgets_values":[696289942349438,"randomize",30,4,"euler_ancestral","simple",1]},{"id":3,"type":"CLIPLoader","pos":[-228.88950749546544,47.59814417305235],"size":[270,106],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"CLIP名称","name":"clip_name","type":"COMBO","widget":{"name":"clip_name"},"link":null},{"localized_name":"类型","name":"type","type":"COMBO","widget":{"name":"type"},"link":null},{"localized_name":"设备","name":"device","shape":7,"type":"COMBO","widget":{"name":"device"},"link":null}],"outputs":[{"localized_name":"CLIP","name":"CLIP","type":"CLIP","links":[7,18]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"CLIPLoader"},"widgets_values":["qwen_3_06b_base.safetensors","stable_diffusion","default"]},{"id":7,"type":"VAELoader","pos":[874.4843317633563,-149.2704472025076],"size":[270,58],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"vae名称","name":"vae_name","type":"COMBO","widget":{"name":"vae_name"},"link":null}],"outputs":[{"localized_name":"VAE","name":"VAE","type":"VAE","links":[4]}],"properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"VAELoader"},"widgets_values":["qwen_image_vae.safetensors"]},{"id":9,"type":"CLIPTextEncode","pos":[533.6664527049207,26.33534263254541],"size":[400,200],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"clip","name":"clip","type":"CLIP","link":18},{"localized_name":"文本","name":"text","type":"STRING","widget":{"name":"text"},"link":null}],"outputs":[{"localized_name":"条件","name":"CONDITIONING","type":"CONDITIONING","links":[8]}],"title":"正面","properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"CLIPTextEncode"},"widgets_values":[""]},{"id":10,"type":"CLIPTextEncode","pos":[195.55231120318282,332.86893478897343],"size":[400,200],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"clip","name":"clip","type":"CLIP","link":7},{"localized_name":"文本","name":"text","type":"STRING","widget":{"name":"text"},"link":null}],"outputs":[{"localized_name":"条件","name":"CONDITIONING","type":"CONDITIONING","links":[9]}],"title":"反面","properties":{"cnr_id":"comfy-core","ver":"0.11.1","Node name for S&R":"CLIPTextEncode"},"widgets_values":[""]}],"links":[[3,4,0,6,0,"LATENT"],[4,7,0,6,1,"VAE"],[5,6,0,8,0,"IMAGE"],[7,3,0,10,0,"CLIP"],[8,9,0,4,1,"CONDITIONING"],[9,10,0,4,2,"CONDITIONING"],[11,12,0,4,3,"LATENT"],[17,2,0,4,0,"MODEL"],[18,3,0,9,0,"CLIP"]],"groups":[],"config":{},"extra":{"workflowRendererVersion":"LG","ds":{"scale":0.984973267580908,"offset":[397.79582369815535,432.0724594528042]}},"version":0.4}

加载后，在左侧找到工作流选项

按住拖到右侧画布里

确认节点都正常显示，再检查三个关键文件有没有选对：

UNET：anima-preview2.safetensors
CLIP：qwen_3_06b_base.safetensors
VAE：qwen_image_vae.safetensors

如果节点红了，一般是模型文件没放对位置，或者文件名和工作流里写的不一致。重新选择一次即可。

提示词写法

anima 可以用 tag，也可以用自然语言。

tag 写法示例：

Frieren, long white hair, green eyes, white dress, fantasy forest, soft lighting

自然语言写法示例：

A quiet fantasy portrait of Frieren standing in a misty forest, wearing a white dress, soft morning light, delicate expression, cinematic composition.

复杂场景更提议用自然语言，把动作关系说清楚：

First-person point of view. The viewer is holding a young woman in a princess carry. She leans into the viewer's chest, looking shy and relaxed. Warm indoor light, soft anime style, intimate composition.

如果模型开始乱画，可以把关系拆得更明确一点。少写抽象词，多写谁在做什么、镜头在哪里、主体占画面什么位置。

作者推荐设置

作者在 Hugging Face 页面里写了推荐设置和使用方法。

打开：

https://huggingface.co/circlestone-labs/Anima

往下滑到 README，或者直接点击页面里的 README 查看即可。

提议第一次使用时先照着作者推荐参数跑一张，确认环境没问题，再改尺寸和提示词。

版本提醒

补充一点：提议使用 10.0 以上版本。界面里可以点击切换版本。

我自己使用的版本没有报错。其他版本不敢保证完全一致。一般情况下，就算出现问题也多半只是 warning，大致率和秋叶整合包里的插件版本支持有关。

如果遇到无法运行，优先检查这几项：

ComfyUI 版本是否太旧
模型文件是否放对目录
UNET / CLIP / VAE 是否选对
秋叶整合包插件是否需要更新
工作流里的文件名是否和本地一致

anima-preview2 最吸引人的地方，就是它对自然语言的理解明显更舒服。许多用 tag 很难讲清楚的姿势、视角、动作关系，用普通句子反而更容易表达。

之后就可以慢慢试自然语言提示词，把那些以前很难描述清楚的画面一点点调出来。

来源地址：

https://www.cocoloop.cn/t/topic/6346

如果你对AI 生图、vibe coding、AI白嫖教程一类的内容感兴趣，也欢迎来CocoLoop逛逛。会持续整理教程、工作流、模型体验、踩坑记录，也欢迎大家发自己的作品和玩法。

社区地址：
https://www.cocoloop.cn/?bbs

欢迎来发帖，来提问，来分享！