v0.5.1 找地面不误判、脚掌贴地、深度图更省

更新方法

• 插件：插件仓库自动更新到 0.5.1。
• 后端（一键包）：设置页点「更新后端」。会照样检查面捕、深度图组件，缺的自动补齐（从 0.4.0 及更早版本升级的会下载约 64MB + 400MB；0.5.0 已经装好的不用重下）。下载不了可以看 手动更新后端。
• 已识别过的图读的是缓存，点工具栏的「重新识别当前图片」才会用上新的找地面和深度图。

新内容

• 找地面更准：只在看得见脚时找地面，并检查人确实站在那个面上。半身照、特写不会再把背景墙、衣服当成地面，人不会被扳歪（之前可能歪 17~68°）；俯拍等真有地面的照片照样扶正
• 脚掌贴地：修正了脚的静止姿态对应，站着的人不再脚尖翘起、只有脚后跟着地
• 新开关「踩地的脚放平」（调整页 → 站姿，默认开）：踩在地上的脚换成游戏里平放的样子，只转脚、不动腿；踮脚、勾脚这类明显故意的不会动
• 深度图更省：分辨率调到 1200 块，结果基本不变（30 张照片对比，地面判断完全一致），显存峰值从约 1.4GB 降到约 0.75GB，CPU 模式从约 3 秒降到约 1 秒
• 后端更新照样会检查并补齐面捕、深度图组件；从 0.4.0 及更早版本升级的，在设置页点「更新后端」一次装好
• 改了识别相关的设置：已识别过的图片读的是缓存，点工具栏的「重新识别当前图片」按钮才会用上新的找地面和深度图

---

How to update

• Plugin: updates to 0.5.1 automatically through the plugin repository.
• Backend (one-click package): click "Update backend" on Settings. It still checks for the face capture and depth components and adds any that are missing (about 64 MB + 400 MB when coming from 0.4.0 or earlier; 0.5.0 users already have them). If downloads fail, see Updating the backend by hand.
• Recognized images come from the cache: use the toolbar's "Recognize the current image again" button to apply the new floor detection and depth settings.

What's new

• Better floor detection: the floor is only searched for when the feet are visible, and the person must actually stand on it. Half-body shots and close-ups no longer mistake a background wall or clothing for the floor and get tilted (previously by 17–68°); photos with a real floor, such as high-angle shots, are still straightened
• Feet flat on the ground: the feet's rest pose is now matched correctly, so standing characters no longer tilt their toes up and touch the ground with the heel only
• New "Flatten planted feet" option (Tuning → Stance, on by default): feet standing on the ground are laid flat like in game, turning only the foot and leaving the legs alone; clearly deliberate tiptoe or flexed feet are kept
• Lighter depth: the depth model runs at 1200 tokens with practically the same results (identical floor decisions on 30 test photos); peak VRAM drops from about 1.4 GB to about 0.75 GB, CPU mode from about 3 s to about 1 s
• Backend updates still check for and add the face capture and depth components; updating from 0.4.0 or earlier needs one click on "Update backend" in Settings
• Recognized images come from the cache: use the toolbar's "Recognize the current image again" button to apply the new floor detection and depth settings

v0.5.0 面部表情、灯光、构图截图与更快的识别

更新方法

• 插件：插件仓库自动更新到 0.5.0。
• 后端（一键包）：设置页点「更新后端」。这次会自动补齐新功能要的组件（面捕约 64MB、深度图约 400MB），不用重新下载一键包；下载慢可以填「下载加速前缀」，断了再点一次，已下好的组件不用重下。
• 面部表情第一次用要在游戏里输入 /posegraph facedump 导出一次游戏表情数据。

新内容

• 后端更新会自动补齐新功能要的组件（面捕约 64 MB、深度图约 400 MB）：老一键包在设置页点一次「更新后端」就行，不用重新下载整个一键包。下载慢可以填下载加速前缀，断了重试时已下好的组件不用重下
• 面部表情（测试）：照片里的表情换算成游戏的脸部骨骼——闭眼、眯眼、挑眉、嘴角、张嘴、歪嘴、视线都按照片的实际幅度还原，最多到游戏里这块脸能动的极限。第一次用要在游戏里输入 /posegraph facedump 导出一次游戏表情数据（一两分钟）
• 分项权重：眼睑、眉毛、嘴部、下巴、视线可以分开调，哪一块不像就单独调哪一块（调整页「面部表情」里，默认展开）
• 头部朝向跟脸：用面捕修正头的朝向，低头、侧脸的角度更准（默认一半，调整页「面部表情」里的「头部跟脸」滑条，不开表情也生效）
• Brio 改动收进图库：应用姿势后在 Brio 里手动改了骨骼，主窗口会提示「检测到在 Brio 里改了 N 根骨骼」，点「收进修正并存进图库」就变成这张图的修正，下次打开自动套上，改别的调整值重新应用时也不会被冲掉
• 灯光：工具栏灯泡按钮，一键给当前角色打光（11 套方案），灯跟着脸和镜头摆；独立的 3D 灯光视图里可以拖动、旋转每盏灯，单独看某一盏或关掉某一盏
• 构图与截图：构图框（横竖构图、遮暗框外、三分线等参考线）、等效全画幅焦距、按原图比例一键适配，可叠加原图做参考；一键截图直接存进 Posegraph
• 复刻镜头更稳：按帧等镜头稳定，低帧率时不再偶尔复刻歪；宽幅照片的视场角更准
• 深度图（MoGe-2）：识别时顺带估计镜头焦距和地面，人物站位、镜头复刻更准；设置页「深度辅助识别」可以关掉（关掉更快、更省显存）
• 识别更快：移植 Fast SAM 3D Body，单人识别约快一倍（GPU 约 0.6 秒，设置页「快速识别」）
• 3D 页：骨骼旋转环（像 Brio 一样选轴直接拖）；右键 / 中键拖动平移视图
• 调整页新增「整体旋转 / 位移」：整个人歪了、要抬高放低时用
• 并行识别（实验）：批量识别时按显存余量开几路同时识别，失败自动退回逐张
• 修复：图库图片多时缩略图可能把游戏冲崩；连本机后端不再走系统代理；后端解码不了的图片改用 Windows 解码重试；3D 页底部勾选框被切掉一截、英文界面下拉框文字被截断
• 合并模式（一张图多人）还在打磨，这一版先关着，下个版本开放

---

How to update

• Plugin: updates to 0.5.0 automatically through the plugin repository.
• Backend (one-click package): click "Update backend" on Settings. This time it also adds the components new features need (face capture about 64 MB, depth about 400 MB), so you don't need to download the package again. Set a download prefix if it's slow; if a download breaks, click again and finished components aren't downloaded twice.
• For facial expressions, type /posegraph facedump in game once to export the game's expression data.

What's new

• Backend updates now add the components new features need (face capture about 64 MB, depth about 400 MB): older one-click packages just click "Update backend" once on Settings, no need to download the whole package again. Set a download prefix if it's slow; finished components aren't downloaded again when you retry
• Facial expressions (beta): the photo's expression is turned into the game's face bones. Eyes closing or squinting, brows, mouth corners, open or crooked mouths and gaze follow the photo's actual amount, up to what each part of the face can move in game. The first time, type /posegraph facedump in game to export the game's expression data once (a minute or two)
• Per-part weights: eyelids, brows, mouth, jaw and gaze can be tuned separately, so you only adjust the part that looks off (under Facial expressions on the Tuning tab, expanded by default)
• Head follows face: face capture corrects the head direction, so tilted and turned heads are more accurate (half by default; the "Head follows face" slider under Facial expressions, works even with expressions off)
• Keep Brio edits in the library: if you change bones in Brio after applying, the main window says how many bones you changed; "Keep as correction and save to library" stores them with the image, so they come back next time and re-applying after other tweaks no longer wipes them
• Lighting: the light bulb on the toolbar lights the current character in one click (11 setups) and keeps the lights placed relative to the face and camera; a separate 3D lighting view lets you drag and rotate each light, solo it or mute it
• Framing & screenshots: a framing box (landscape or portrait, dim outside, thirds and other guides), equivalent full-frame focal length, one-click fit to the photo's aspect, an optional photo overlay for reference, and one-click screenshots saved into Posegraph
• Steadier camera matching: waits for the camera by frames, so low frame rates no longer occasionally match it crooked; wide photos get a more accurate field of view
• Depth (MoGe-2): recognition also estimates the camera focal length and the floor, making placement and camera matching more accurate; turn off "Depth-assisted recognition" on Settings for speed and less VRAM
• Faster recognition: ported from Fast SAM 3D Body, single-person recognition is about twice as fast (about 0.6 s on GPU, "Fast recognition" on Settings)
• 3D tab: rotation rings on bones (pick an axis and drag, like Brio); pan the view with the right or middle mouse button
• Tuning tab: new "Whole-body rotation / offset" for when the whole person leans or needs raising or lowering
• Parallel recognition (experimental): batch recognition runs several at once based on free VRAM and falls back to one by one on failure
• Fixes: library thumbnails could crash the game with many images; the local backend no longer goes through the system proxy; images the backend can't decode are retried via Windows; the 3D tab's bottom checkbox was cut off and dropdown text was truncated in English
• Merge mode (several people in one image) is still being polished and stays off in this version; it opens in the next one

后端组件 1 / Backend components 1

插件内「更新后端」自动下载用的组件，不用手动下载。老一键包缺哪个补哪个（见 backend/src/posegraph/components.json）。

• posegraph-face-1.zip：面捕（MediaPipe Face Landmarker，Apache-2.0）
• posegraph-depth-1.zip：深度图（MoGe-2 ViT-B，MIT）

Components downloaded automatically by "Update backend" in the plugin; no need to download them by hand. Older one-click packages get whichever is missing (see backend/src/posegraph/components.json).

• posegraph-face-1.zip: face capture (MediaPipe Face Landmarker, Apache-2.0)
• posegraph-depth-1.zip: depth (MoGe-2 ViT-B, MIT)

Posegraph 0.4.0

0.4.0 拆分模式、辅助视角与新界面

更新插件后，请在插件「设置」页更新后端（一键包用户点「更新后端」即可，依赖没有变化）。

• 角色列表（拆分模式）：GPose 里的多个角色各绑一张图（从图库拖进来），一键全部应用；某个角色要细调时点「编辑」切到主窗口
• 辅助视角：给一张图挂上同一动作其他角度的图（最好是约 90° 的侧面），自动融合修正单张图猜不准的前后深度；主图里看到的姿势不会被改动。在图库右键「辅助视角」里管理，文件名为「主图名-1」的图导入时自动挂上
• 批量识别只加载一次模型：主图连同辅助视角、全部应用、批量队列都更快（“识别后立即释放”时尤其明显）
• 调整页的数值随「保存修正」按图片保存；新图片都从默认值开始，不再沿用上一张图调过的值
• 全新界面：深灰 + 青蓝主题，工具栏分三行，默认窗口大小下不再裁切；放不下的文字以「…」结尾、悬停显示全文
• 更新日志窗口：更新后第一次打开主窗口时弹出，设置页可以再打开

0.3.9 / 0.3.10（之前未单独发布）

• 图库：导入（多选、文件夹、拖放、剪切板）、文件夹 / 标签 / 收藏分类、重命名、移动、删除到回收站；识别结果按图片内容缓存，再次打开不用重新识别；批量识别；存到 Brio 姿势库
• 3D 页「保存修正」：手动偏移跟着图片走
• 女性骨架适配；膝、肘、肩辅助骨规则按游戏内采样更新
• 切换角色不再重置该角色在 Brio 里的姿势；角色列表只列人形角色；检测 IVCS
• 修复多次换种族后骨长错乱
• 许可证改为 MIT

---

0.4.0 Split mode, extra views and a new look

After updating the plugin, update the backend from the plugin's Settings tab (backend package users: click "Update backend"; dependencies are unchanged).

• Character list (split mode): bind one image to each GPose character (drag from the library) and apply them all at once; click Edit to fine-tune one in the main window
• Extra views: attach images of the same pose from other angles (a ~90° side view is best). They are fused automatically to fix the front/back depth a single image can't tell, without changing the pose you see in the main image. Manage them from the library's right-click menu; images named "main-1" are attached automatically on import
• Batch recognition loads the model only once: main image plus extra views, Apply all and the batch queue are all faster
• Tuning values are saved per image with "Save correction"; new images start from the defaults
• New look: dark gray with a teal accent, three-row toolbar that no longer gets cut off; text that doesn't fit ends with "…" and shows in full on hover
• Changelog window: pops up the first time you open the main window after an update; reopen it from Settings

0.3.9 / 0.3.10 (not released separately)

• Image library: import (multi-select, folders, drag & drop, clipboard), folders / tags / favorites, rename, move, delete to the Recycle Bin; recognition results cached by image content; batch recognition; save to the Brio pose library
• "Save correction" on the 3D tab: manual offsets travel with the image
• Female skeleton support; knee, elbow and shoulder helper bone rules updated from in-game sampling
• Switching characters no longer resets their Brio pose; the character list only shows humanoids; IVCS detection
• Fixed bone lengths going wrong after changing race several times
• License changed to MIT

Posegraph 0.3.7

Fixes

• The window did not show in GPose (not even with /posegraph) when Dalamud's hide plugin UI in GPose setting was on. Posegraph now always shows in GPose, like Brio.
• New option Show while the game UI is hidden (off by default, like Brio). Turn it on if the window disappears when you hide the UI.
• A window that opened automatically on entering GPose now closes when you leave GPose. A window opened with /posegraph or from the plugin list stays open.

Package users: update the backend from Settings (version number only).

---

修复

• Dalamud 开启「GPose 中隐藏插件界面」时，进入 GPose 窗口不显示，输入 /posegraph 也不显示。现在 GPose 中始终显示，与 Brio 一致。
• 新增「游戏界面隐藏时仍显示」选项（默认关，与 Brio 一致）。隐藏游戏界面后窗口消失时可以打开。
• 进入 GPose 自动打开的窗口，退出 GPose 时自动关闭；用 /posegraph 或插件列表打开的保持不变。

一键包用户：在设置页更新后端（只有版本号变化）。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)

Posegraph 0.3.6

Reverts the 0.3.5 camera matching changes; behaviour is the same as 0.3.4. Update the backend in Settings.

---

撤回 0.3.5 的镜头复刻改动，行为与 0.3.4 相同。请在设置页更新后端。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)

Posegraph 0.3.5

Camera matching

• Keeps the actor in frame. Body parts visible in the photo (top of the head, fingers, toes, …) must also be on screen in game. If they are not (tight photo framing, unreachable camera position, different window shape), the view is widened first; once the view is at its limit, the camera moves straight back. Can be turned off in Settings.
• Landscape windows. The view always contains the whole photo frame for the game window's aspect ratio (16:9, 16:10, 21:9, …): matched by height when the window is wider than the photo, by width when it is narrower.
• The framing error is now measured in view angles, independent of window shape and widened views.

Backend update required (Settings → download button next to "Backend …, plugin …").

---

镜头复刻

• 保证角色完整入画。 照片里拍到的部位（头顶、手指、脚尖等）在游戏里也必须在画面内。出画时（照片构图很满、机位够不到、窗口比例不同），先拉宽视角，视角到上限后再沿视线后退。设置里可关闭。
• 适配横屏窗口。 视角按游戏窗口比例（16:9、16:10、21:9 等）完整包含照片画面：窗口比照片宽时按高度对齐，比照片窄时按宽度对齐。
• 构图偏差改为按视线角度计算，与窗口比例、拉宽的视角无关。

需要更新后端（设置页「后端版本 …，插件 …」旁的下载按钮）。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)

Posegraph 0.3.4

Change

• Tuning values (amplitudes, hand compensation, posture, helper bones) are saved per preset (Photo / Illustration) and restored next time. Manual joint offsets are per image and not saved. The reset button next to the preset restores the defaults.

Package users: update the backend from Settings (only the version number changed).

---

改动

• 调整页的值（幅度、手部末端补偿、站姿、辅助骨）按预设（照片 / 插画）自动保存，下次打开时恢复。手动关节偏移只对当前图有效，不保存。预设旁的恢复按钮会恢复默认值。

一键包用户：在设置页更新后端（只有版本号变化）。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)

Posegraph 0.3.3

Fix

• After a Glamourer race/appearance change, the pose is recalculated and applied automatically if it was applied to that actor before (0.3.2 only recalculated).
• Waits until the model has finished reloading before reading the new bone lengths.

Package users: update the backend from Settings (only the version number changed).

---

修复

• 用 Glamourer 换种族或外貌后，如果这个角色之前应用过姿势，重新计算后会自动应用（0.3.2 只重新计算）。
• 等模型重建完成后再读取新的骨长。

一键包用户：在设置页更新后端（只有版本号变化）。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)

Posegraph 0.3.2

Fixes

• Wrong body proportions after changing race (Lalafell "stick insect", shrunken Roegadyn). The game pose that supplies the real bone lengths was cached per actor only, so after a Glamourer race change the old race's lengths were reused. It is now keyed by actor + model race + appearance and re-read automatically when any of them changes.
• Race detection reads the model's race, so Glamourer changes are picked up.
• The backend warns when the game bone lengths do not match the selected skeleton.

Changes

• The main window opens automatically when entering GPose with Brio ready (can be turned off in Settings).
• Toolbar dropdowns are sized to their content.
• Midlander-type and Elezen males are marked as supported; the "untested" label is removed.

Package users: update the backend from Settings (download button next to "Backend …, plugin …").

---

修复

• 换种族后体型比例错误（拉拉菲尔变"竹节虫"、鲁加系缩成一团）。 提供游戏内骨长的参考姿势之前只按角色缓存，Glamourer 换种族后仍用旧种族的骨长。现在按角色 + 模型种族 + 外貌缓存，任一变化都会自动重新读取。
• 种族识别改为读取模型种族，能跟随 Glamourer 的修改。
• 游戏骨长与所选骨架不符时，后端会提示。

改动

• 进入 GPose 且 Brio 就绪时自动打开主界面（设置里可关闭）。
• 工具栏下拉框按内容加宽。
• 中原系男、精灵男标为已适配，去掉「未实测」标记。

一键包用户：在设置页点「后端版本 …，插件 …」旁的下载按钮更新后端。

Backend package / 后端一键包: Posegraph-Backend-win64.zip (Google Drive)