1000+ 开源白模如何炼成:EverOS 终极实战

CC0 开源白模姿势参考库项目头图

21 小时,3 个主力 Agent 协同,我从零造了一个 CC0 开源白模五机位姿势参考库;工具:EverOS + Claude Code + Muse + Grok Bot

前两天几十万浏览的那个姿势库镜像,原作条款其实写了禁止再分发,我干脆用 21 小时自己重做了一套 CC0 免费版,把图片从接近 1 GB 压缩到 50 MB 左右,可自由商用、二次分发、二次创作。(🔗 见最后)

姿势库的成品预览

本文内容:

  1. 多 Agent 协作流程(适合关注基于记忆的多 Agent 协作的朋友)
  2. 9 个关键提示词全拆解(适合关注完整提示词的朋友)
  3. 成品与开源地址(推荐 ✨)

16 小时

昨天几十万浏览的那个姿势库镜像,原作条款其实写了禁止再分发,我干脆用 21 小时自己重做了一套 CC0 免费版。一共 288 个姿势 × 5 个机位,1,133 张质检过的线稿,全靠 3 个 AI agent 用共享记忆接力跑完。拿到白模后配张角色图喂给生图模型就能换姿势,工作流随后放出。

现在网站上有 1,133+ 张质检过的线稿和 89+ 张实拍对照,覆盖 288 个姿势 × 5 机位(正面、斜侧、侧面、俯视、仰视),支持中、日、英、韩四语,图片和数据采用 CC0 协议。

EverOS 实战演示视频封面

1. 多 Agent 协作流程

整个项目里,我把 EverOS(跨应用共享记忆层)当成几个独立工具之间的传话筒,工作流分工如下:

多 Agent 协作流程图

  1. Grok Bot 前期调研 排查原图版权红线、图片压缩方案和 4 种生图路线,聊完把结论存进 EverOS。
  2. 1 分钟跨工具交接 我切到 Muse 新开窗口,第一句话就从 EverOS 读到了 1 分钟前刚存进去的调研记忆,几分钟后本地的 Claude Code 也一样接上了。
  3. 双线并行生图 Muse 自带生图模型交出主力线稿;同时一个 Claude Code 会话在 compshare 上租了台 RTX 4090,跑 ComfyUI(Qwen-Image-Edit-2511 fp8 + Lightning 4 步 LoRA)专门补画缺漏。
  4. 多会话协同质检与建站 十几个本地 Claude Code 会话同时开工,派 7 个 Sonnet 子代理并行审图、写前端、做部署。
  5. 经验写回 收工前,Muse 把整套流程分 9 块写回 EverOS 并沉淀成一个 skill,Claude Code 也把 ComfyUI 参数和避坑记录存了进去。

三个 Agent 接的都是托管版 EverOS MCP,同一个账号登录就是同一份记忆。
想复现的话,把下面这段发给你手头的 Agent 就行:

帮我安装 EverOS 的 MCP,用托管版 https://everos-mcp.evermind.ai/mcp ,走 OAuth 登录我的 EverOS 账号。装好后调一次 get_profile 确认连通。以后我说“存到 everos”就用 add_memory 写入,说“从 everos 查”就用 search_memory。

跨工具交接时,我对 Grok Bot 说的最后一句话是:

帮我保存一下记忆到 everos 吧

1 分钟后,我在 Muse 新开窗口说的第一句话是:

你从 everos 能获取姿势库的记忆信息吗

跨工具共享记忆的对话记录

2. 个关键提示词全拆解

整套图库最核心的部分,是怎么把控无脸白模的画风、人体体块和五机位透视。
下面按实际生产顺序贴出关键提示词。

2.1. 姿势描述词(288 个姿势的动作底座)

无论是用 Muse 还是 ComfyUI,每个姿势都需要一段准确的英文动作描述。
现在网站上每个姿势点开都附带了这段可复制的词,以 P001 为例:

sitting on the floor leaning back with her torso arched slightly backward, one knee raised and bent, the other leg extended forward, both hands resting on the floor behind her for support

2.2. Muse 量产管线(写实图转线稿)

我对照原作拆解出一条稳妥的路线,先生成动作准确的写实图,再用一段固定模板转成线稿([具体姿势描述] 处填入上面的动作词):

faceless mannequin, pose reference, line art, black line only, monochrome, white background, no face, no hair, no clothes, joint circles, clear body structure, [具体姿势描述], full body, 3:4

为了让 Agent 看图时能抓准胸腰体块,光在提示词里硬加词会把风格带偏,导致关节圆变成实心黑点。
Muse 试出来要两个杠杆一起用,先让写实图里的人物穿上修身衣服(白背心 + 黑 leggings),转线稿时再往上面的模板后面追加这一句:

feminine female body with a clearly defined bust contour in the outline and a slender waist

后来在全量跑时,为了绕开模型把写实人体转无衣素体时频繁触发的安全拒绝,Muse 的最终定稿模板在上面模板的基础上去掉了照片那一步,直接按姿势描述纯文字出线稿。

2.3. ComfyUI 云端补缺管线(锁体型、转机位、转线稿)

Claude Code 在 4090 上跑的开源模型管线拆得更细。
第一步是先生成一张标准女性素体渲染图,后续所有姿势都从这张图出发去改,用来锁死体型比例:

A photorealistic 3D render of a faceless adult female art mannequin standing straight in a relaxed A-pose, arms slightly away from the body, feet slightly apart, full body, front view. Curvy figure: full rounded chest, narrow waist, wide rounded hips and buttocks, toned thighs. Smooth seamless matte light grey plastic skin like a store display mannequin; the chest is completely smooth like a Barbie doll, with no nipples; no clothing. Smooth egg-shaped head without face or hair. Plain white studio background, soft even lighting.

第二步,把上面这张基准素体作为 image 1 垫进去,只改姿势。
因为用了 Lightning 加速 LoRA(cfg=1),负向提示词不起作用,所以“不要椅子道具、不露点”等忌讳全部要写成正向描述(以 P026 为例):

Change only the pose of the mannequin in image 1. Keep exactly the same mannequin: the same body proportions, the same smooth seamless matte light grey plastic skin, the same completely smooth chest without nipples, the same blank egg-shaped head without face or hair, no clothing, plain white studio background, soft even lighting. Nothing else in the scene: no chair, stool, box, bed or any furniture or props; any seated pose sits directly on the plain white floor. New pose: standing with one arm raised up, her hip pushed out to one side in a playful pose Framing: full body, all limbs fully inside the frame. Front view, eye-level camera.

第三步,拿到正面写实图后,用多角度 LoRA 生成其余 4 个机位的写实图(约 7 秒一张)。
早期写 60 度俯视和正面仰视会出现照抄或机位重复,最后定稿改成 30 度俯视和左前 45 度仰视,每个机位对应一句触发词:

diag:  <sks> front-right quarter view eye-level shot wide shot
side:  <sks> right side view eye-level shot wide shot
above: <sks> front view elevated shot wide shot
below: <sks> front-left quarter view low-angle shot wide shot

第四步,双图参考转线稿(约 12 秒一张)。
image 1 放刚生成的写实姿势图,image 2 放画风参考模板。
这里踩过一个大坑,如果 image 2 给整身模板,模型就会无视提示词直接照抄模板的站姿;把模板裁剪到“头部至大腿根”之后,照抄率降到了 0。
以 P026 斜侧机位为例:

Redraw the figure in image 1 as a pose reference drawing in exactly the drawing style of image 2. Faceless mannequin, pose reference, line art, bold black outer contours with thin light grey construction lines, white background, no face, no hair, no clothes, joint circles, clear body structure. The face area stays completely blank: no eyes, eyebrows, nose, mouth or lips. Draw only the mannequin itself: no furniture, props, floor, walls, shadows, guide lines or background of any kind. Copy the body construction of image 2: blank egg-shaped head with a thin vertical center line and a curved eye line, a center line down the torso, breasts drawn as simple smooth empty circles with nothing inside them (no nipples, no dots), a curved line under the chest marking the ribcage, a curved pelvis band, thin axis lines along the arms and legs, and large round ball joints at the shoulders, elbows, wrists, knees and ankles. The drawing shows a curvy female mannequin with a full chest and prominent rounded hips and buttocks. Keep exactly the pose, camera angle and framing of the figure in image 1, never the pose or arm position of image 2 (image 2 is only a style sample). Camera: three-quarter view from the front-right exactly as in image 1; the body is turned, not facing the viewer. Pose: standing with one arm raised up, her hip pushed out to one side in a playful pose Framing: full body, all limbs fully inside the frame, 3:4.

在上面第四步里,如果只写“保持 image 1 的机位”还不够,俯拍坐姿经常会被模型强行画回正面站姿。
所以必须把 Camera: 那一句按 5 个机位分别写死替换进去:

front: Camera: front view at eye level.
high: Camera: seen from above, looking down at the figure, with strong top-down foreshortening exactly as in image 1 (head and shoulders large and close, legs short and far); never redraw it as a standing front view.
low: Camera: low angle from the front-left, looking up at the figure, turned about 45 degrees exactly as in image 1.
side: Camera: right side profile view exactly as in image 1; the body is seen from the side, not from the front.
diag: Camera: three-quarter view from the front-right exactly as in image 1; the body is turned, not facing the viewer.

2.4. Sonnet 子代理质检(肢体计数专项)

常规质检跑完后,我肉眼抽查发现 P006 正面画了三只手。
我让 Claude Code 专门补了一轮肢体计数检查,发给并行审图的 Sonnet 子代理,要求顺着肩膀和胯部一节节往下数,这一轮又拦下了 41 张多手多脚的图:

For each panel, before deciding anything:
1. Find each shoulder. Follow every arm from its shoulder ball joint through the elbow and wrist to its hand. Write
   down where each hand ends up (e.g. "R hand on hip, L hand on floor behind").
2. Find each hip. Follow every leg from the hip through the knee and ankle to its foot.
3. Then look for any arm, forearm, hand, leg or foot you did NOT reach from a shoulder or hip — including one
   partly hidden behind the body or another limb, or a second limb leaving the same shoulder or hip. Each of those is
   an extra limb. A hand or foot drawn twice (two hands on one wrist) also counts.
4. Count: arms, hands, legs, feet. One figure must give 2 / 2 / 2 / 2; "duo" (two figures) 4 / 4 / 4 / 4.

3. 成品与开源地址

白模线稿的成品效果

多机位成品对照

你平时把手头的多个 AI 工具串起来干活时,是怎么解决它们之间上下文同步问题的?
欢迎在评论区聊聊你的工作流,如果去试用了 Shisei 图库,也告诉我你还缺哪些姿势。

相关链接:

  1. EverOS:https://everos.evermind.ai/(免费额度很够,现在订阅 Pro 还能每月额度升到 3 倍,连享 6 个月)
  2. Shisei 官网(1000+ CC0 协议白模):https://shisei.indaloop.app/
  3. 代码与数据全部开源在 GitHub 仓库:https://github.com/wquguru/shisei