需要 Android 10 及以上手机,并先安装 Hermit。参考图由手机现场渲染,生图模型可以由你自己的电脑或云端服务提供。Needs Android 10 or later plus Hermit. The phone renders the reference itself, and the image model can come from your own computer or a cloud service.
PoseGi3D 视口viewport右手已选中Right hand picked
站在礁石上回头,风把外套吹开Turning back on the rocks, wind pulling the coat open参考图 1024Reference 1024
界面示意:一根人形骨架站在视口里,选中右手后拖动,肩膀与手臂会一起让开。Interface sketch: a humanoid skeleton in the viewport; pick the right hand and drag, and the shoulder and arm give way with it.
摆姿Posing
拖一个关节, 整条链都会动。Drag one joint. The whole chain follows.
PoseGi 不让你输入一串角度数字,也不让你正对着滑杆猜。骨架摆在那里,点中一个关节,直接拖。拖的时候两颗约束同时生效:每个关节只能在自己的解剖学范围内转,而 IK 会把改动顺着骨链传下去,让手腕而不是整条胳膊看起来断掉。PoseGi does not ask you to type angles or guess at sliders. The skeleton is right there: tap a joint and drag it. Two constraints apply while you drag. Each joint only turns inside its own anatomical range, and the IK carries the change down the bone chain so the wrist does not snap while the arm goes with it.
关节表是纯数据:骨盆,腰,胸,颈,头,加上左右各一套肩,上臂,前臂,手,大腿,小腿,脚。The joint table is plain data: hips, spine, chest, neck and head, plus shoulders, upper arms, forearms, hands, thighs, shins and feet on both sides.
每个关节都带自己的可转范围,比如手腕 ±80°,脚踝 -50°..25°,膝盖 -150°..0°。写入前统一收口,越界直接被夹住。Every joint carries its own range: a wrist at ±80°, an ankle from -50° to 25°, a knee from -150° to 0°. All writes go through one clamp, so nothing goes past it.
拖「连接杆」是旋转,拖「节点」是移动 —— 同一个手势,不用先切模式。Dragging the bone rotates, dragging the node moves. One gesture, no mode to switch first.
绕骨轴自转是拖拽永远碰不到的自由度(转轴就在骨头自己身上),所以它单独有一根滑杆,行程就是该关节自己的范围。Twisting around the bone's own axis can never be reached by dragging, because the axis runs through the bone itself. So it gets its own slider, with the joint's own range as its travel.
骨架与它的边界The rig and its limits
躯干Trunk
骨盆 / 腰 / 胸 / 颈 / 头Hips, spine, chest, neck, head
骨盆可以绕垂直轴整圈转,腰与胸负责前倾后仰和侧弯,颈与头单独微调朝向。重心动一下,整个站姿就不一样了。The hips turn a full circle around the vertical axis, spine and chest handle leaning and bending, and neck and head aim on their own. Move the weight and the whole stance changes.
右臂Right arm
肩 / 上臂 / 前臂 / 手Shoulder, upper arm, forearm, hand
肩有前伸与上举两个方向,上臂活动度最大,肘只做屈伸,手腕再补一段 ±80° 的精细调整。拖动时 IK 会让整条链一起让开。The shoulder reaches forward and lifts, the upper arm has the widest travel, the elbow only flexes, and the wrist adds a fine ±80°. Drag any of them and the IK opens the whole chain.
双腿Both legs
大腿 / 小腿 / 脚Thigh, shin, foot
髋部前后摆幅大,膝盖只向后弯,脚踝在 -50°..25° 之间找平。两节链用的是闭式解,所以抬腿时膝盖一定会弯,不会退化成一根直杆。The hip swings widely fore and aft, the knee only bends backwards, and the ankle levels out between -50° and 25°. Two-bone chains use a closed-form solve, so a lifted leg always bends at the knee instead of collapsing into a straight rod.
出厂带一个造型(宜家人偶),骨架参数就是它的模型数据。这一层单独留着,是为了以后换第二个人偶时只需要登记一次。One figure ships in the box (the IKEA mannequin) and its model data is what the rig reads. That layer is kept separate so a second figure would only need to be registered once.
界面按钮The interface
按钮就在那儿,一按就会。The buttons are right there. Press one.
下面每个图标,都是应用里真实存在的按钮。视口占满整个屏幕,所有工具都收在底部的按钮条里 —— 3D 场景不跟面板抢地方,取景时也不会有人物的头被面板压住。Every icon below is a real button in the app. The viewport takes the whole screen and every tool lives in the dock at the bottom, so the 3D scene never fights a panel for room and no head gets covered while you frame a shot.
底部常驻按钮The bottom dock
姿态Pose
打开姿态面板:站姿,T 字,行走,举手,坐下,以及镜像与复位。Opens the pose panel: stand, T pose, walk, wave, sit, plus mirror and reset.
选取Pick
点中一个关节。选中的是「连接杆」还是「节点」,决定了接下来拖动是旋转还是移动。Taps a joint. Whether you picked the bone or the node decides if the next drag rotates or moves.
工具Tools
视口工具与取景:调整相机距离与构图,让人物在画幅里站好位置。Viewport tools and framing: set the camera distance and compose the figure inside the frame.
渲染Render
渲染当前造型。这张图既是手机上的预览,也是交给模型的参考图。Renders the current pose. That image is both the preview on the phone and the reference handed to the model.
截图Shot
把当前造型存成一张图,不发起任何模型请求。Saves the current pose as an image without calling any model.
生成Generate
写下画面描述,把现场渲染的参考图交给当前模型卡。生成中再点一次就是忽略这次结果。Write what the picture should be and hand the freshly rendered reference to the active model card. Pressing it again while a job runs ignores that result.
关节与姿态Joints and poses
拖连接杆Drag the bone
旋转这个关节,IK 会把改动传给下游的骨节。Rotates the joint, and the IK passes the change down the chain.
拖节点Drag the node
移动这个关节本身,用来微调整体位置。Moves the joint itself, for nudging the overall placement.
自转滑杆Twist slider
单独解决绕骨轴的自转,两端就是该关节的解剖边界。Handles rotation around the bone axis on its own, with the joint's anatomical limit at either end.
镜像Mirror
左右对调,一步把手上的姿势翻到另一边。Swaps left and right, flipping the pose to the other side in one step.
复位Reset
回到默认站姿,重新开始摆。Back to the default stance, ready to pose again.
作品与成图Artworks and results
添加作品Add artwork
一件作品 = 标题 + 提示词 + 它自己的成图。开新作品就是换一块干净画布。One artwork is a title, a prompt and its own generated images. Starting a new one gives you a clean sheet.
作品列表Artwork list
所有作品与它们的成图都在这里,随时回到任何一件接着改。Every artwork and its images, ready to be picked up again.
全屏预览Full-screen preview
双指缩放,拖动平移,双击复位,看清细节。Pinch to zoom, drag to pan, double-tap to reset, so you can read the detail.
复位图片Reset the image
预览里把缩放和平移打回原样。Puts zoom and pan back where they started inside the preview.
删除成图Delete an image
不要的那一张就地删掉,作品本身还在。Removes the one you do not want while the artwork stays.
参数都摆在明面上Every parameter in the open
参考图强度Reference strength0–200
生成分辨率Output size512–1024 px
参考图边长Reference edge1024 px
步数Steps1–60
超时Timeoutms
自定义请求头Custom headersJSON
翻译为英文Translate to Englishon / off
参考图强度 100 是中性:参考图原样交给模型。往下调会把参考图逐步柔化,让提示词接手;往上调则更紧地贴住骨架。参考图本身永远是 1024 边 —— 换了张 512 的模型卡,模型看到的还是同一张图。A reference strength of 100 is neutral: the model sees the render as it is. Lower softens it so the prompt takes over; higher sticks tighter to the skeleton. The reference itself is always 1024 on its long edge, so a 512 model card still sees the same picture.
模型接入Connecting models
本地,云端,还是自己的协议。Local, cloud, or your own protocol.
PoseGi 不自带 API Key,也不绑定任何平台。它按你建的「模型卡」提交请求,模型和工作流仍然在你手上。公网服务必须 HTTPS,可信局域网可以用明文 HTTP。PoseGi ships no API key and locks you to no platform. It submits requests through the model cards you create, and the models and workflows stay yours. Public services must use HTTPS; a trusted LAN may use plain HTTP.
本地 ComfyUI + VibeDraw 插件节点(推荐)Local ComfyUI with the VibeDraw plugin (recommended)
给 ComfyUI 装上 VibeDraw 插件后,PoseGi 就能把参考图和描述交给它,并按插件自报的能力选任务:快速草图是 512 画幅、一两秒出图;放大任务把 1024 渲染图补细节、构图基本不动;Qwen-Image 2.1 会按参考图重新作画,是这里最重也最像成片的一档,单张几十秒。点「测试连接」时插件会自报它有哪些模型和任务,画面参数随之调整。With the VibeDraw plugin installed, PoseGi hands the reference and the description to ComfyUI and picks a task from what the plugin reports: the quick sketch path is 512 px in a second or two, the upscale task adds detail to the 1024 render without moving the composition, and Qwen-Image 2.1 repaints from the reference, the heaviest and most photographic option at tens of seconds a frame. Test connection asks the plugin what it offers and the form follows along.
任何实现了 OpenAI Images 接口的服务都能直接接:填地址、模型 ID 和 API Key,也可以把参考图一起提交过去。Any service speaking the OpenAI Images API connects directly: an address, a model ID and an API key, and the reference can go along with the request.
SD WebUI / ForgeSD WebUI and Forge
指向你的 WebUI 地址,用图生图把渲染图当作初始图,去噪强度就是参考图强度。Point it at your WebUI and it runs image-to-image with the render as the init image, where the denoising strength is the reference strength.
Stability AIStability AI
直接用官方的 v2beta 接口,把渲染图当作控制图提交,控制强度同样可调。Talks to the official v2beta endpoint and submits the render as the control image, with the control strength just as adjustable.
按你的机器挑任务Pick a task for the machine you have
最轻Lightest
快速草图:512 画幅,一两秒出图。适合先把构图和人物质感确认下来,再决定要不要放大。Quick sketch: 512 px in a second or two, good for settling the composition and the feel before deciding to enlarge.
快速Quick
中等Middle
放大补细节:把手上的 1024 渲染图放大并补细节,构图基本不动。姿势已经对了,只想要更清楚的皮肉。Upscale: enlarges the 1024 render and adds detail while the composition stays put, for when the pose is right and only the surface needs more.
放大Upscale
最重Heaviest
Qwen-Image 2.1 重绘:按参考图重新作画,单张几十秒。画幅 512 到 1024 每 64 一档,中文描述也能直接读。Qwen-Image 2.1 repaint: repaints from the reference at tens of seconds a frame, from 512 to 1024 in steps of 64, and it reads Chinese prompts directly.
Qwen-Image 2.1
基于 Hermit,所以能改Built on Hermit, so it can be changed
整个应用,你都能换成自己的。Change anything. The whole app.
PoseGi 是跑在你自己的 Hermit 里的 MIT 开源应用。在 Hermit 的开发页开启开发模式,把地址和密码交给你信任的智能体,骨架关节表,视口取景,界面配色,生图流程,都能按你的习惯重做 —— 关节范围本身就是一张表,改数据不用改编排。PoseGi is an MIT-licensed app running inside your own Hermit. Turn on development mode from the Hermit development page, hand the address and password to an agent you trust, and the joint table, the framing, the colours and the generation flow can all be rebuilt around your habits. The joint ranges are a table, so changing the behaviour does not mean rewriting the logic.
WorkBuddy桌面智能体Desktop agent或orCodex命令行智能体CLI agent任何支持 MCP 的智能体都可以any MCP-capable agent works
第一步:装 HermitStep one: install Hermit
PoseGi 跑在 Hermit 里,先把它装上。PoseGi runs inside Hermit. Install that first.
Hermit 是手机上装应用的地方,负责安装,数据,文件,权限和系统能力。装好 Hermit,再把 PoseGi 加进去。不接模型也能先摆 —— 骨架,关节,预设,取景和截图都能单独用。Hermit is where apps live on your phone: it owns installation, data, files, permissions and system capabilities. Install Hermit, then add PoseGi. It poses before any model is connected: the rig, the joints, the presets, the framing and the shot all work on their own.
VibeDraw 是同一个 Hermit 里的实时 AI 绘图工具:你在画布上随手涂,成图层跟着变。PoseGi 管的是骨骼和构图,VibeDraw 管的是手感和色彩 —— 一个把人物摆对,一个把画面画美。VibeDraw is the real-time AI drawing tool for the same Hermit: scribble on the canvas and the result layer follows. PoseGi owns the skeleton and the composition, VibeDraw owns the touch and the colour: one gets the figure right, the other makes the picture beautiful.
不存。一件作品 = 标题 + 提示词 + 它自己的成图,姿态不进存储。原因是生成的瞬间手机本来就现场渲染一张参考图,所以没有"先存姿态再生成"的必要。想留住一个造型,点「截图」把它存成图。No. An artwork is a title, a prompt and its own generated images; the pose is not stored. The phone already renders a fresh reference at the moment of generation, so saving first would buy nothing. To keep a pose, use Shot and keep the image.
一定要有独立显卡吗？Do I need a discrete GPU?
只有本地跑模型才有这个要求,而且最轻的快速草图任务只要 512 画幅,门槛不高。不想装环境就用云端服务,手机照样摆姿。Only if you run the model locally, and the lightest quick sketch task runs at 512 px, so the bar is low. If you would rather not set anything up, a cloud service still lets the phone do the posing.
生成的图不太像我摆的姿势？The result does not quite match my pose?
先把参考图强度往上调,让模型更紧地贴住骨架;再检查取景,人物在画幅里太小会让姿势细节被稀释。如果只是想要更清楚的皮肉而不改构图,用「放大」而不是重画。First raise the reference strength so the model sticks closer to the skeleton, then check the framing: a figure that is small in the frame dilutes the pose. If you only want clearer surface without moving the composition, use Upscale rather than a repaint.
本地 ComfyUI + VibeDraw 插件节点(推荐)Local ComfyUI with the VibeDraw plugin (recommended)
