Skip to content

Latest commit

 

History

History
529 lines (389 loc) · 39.1 KB

README.md

File metadata and controls

529 lines (389 loc) · 39.1 KB

Posegraph tutorial

English | 简体中文

This tutorial takes you from installation to putting a photo's pose on your GPose character, then covers the image library, extra views, posing several characters, lighting, video motion and posing right in the game (pose editor, hand snapping). The numbers in each screenshot match the notes on its right. The UI screenshots are made with the offline screenshot tool; the sample photos are credited at the end.

Results: the tutorial's four sample photos recognized and put on an in-game character (Midlander male, with the photo's camera matched):

Leaning on a wall Arms open
Leaning on a wall Arms open
Sitting on a ledge Martial arts leap
Sitting on a ledge Martial arts leap (two people in the photo; only the one in the air is posed here - use merge mode to pose both, see section 10)

Contents

  1. Before you start
  2. Set up the backend
  3. The main window
  4. Recognize your first image and apply it
  5. 3D tab: check and fine-tune joints
  6. Tuning tab: amplitude and corrections
  7. Hints tab
  8. Image library
  9. Extra views: fixing front/back depth
  10. Several characters: the character list
  11. Camera matching, export and restore
  12. Facial expressions (beta)
  13. Keeping Brio edits in the library
  14. Lighting
  15. Framing and screenshots
  16. Video motion (beta)
  17. Convert to idle
  18. Edit in game (pose editor)
  19. Hand snapping and "Place on body part" (experimental)
  20. Updates
  21. FAQ
  22. Community & support

1. Before you start

You need:

  • FFXIV with Dalamud and Brio enabled
  • The Posegraph backend: download Posegraph-Installer.zip, unpack it and double-click Install-Posegraph.bat. Choose the install location (an English-only path such as D:\Posegraph-Backend), compute mode and network, and wait for it to finish (about 3–6 GB). Your GPU is detected and a compute mode recommended. See the installer notes
  • For GPU recognition, an NVIDIA card or a ROCm-capable AMD card (RX 9000, RX 7900 / 7800 / 7700 and others) with about 3 GB of free VRAM; without one it runs on the CPU (about 10–20 s per image)

Install the plugin:

  1. In game, type /xlsettings → Experimental → Custom Plugin Repositories, add https://raw.githubusercontent.com/extrant/DalamudPlugins/main/pluginmaster.json, tick it and save.
  2. Search for Posegraph in the plugin installer and install it.
  3. Type /posegraph to open the main window. The changelog pops up the first time; hold the "Hold to close" button to dismiss it.

2. Set up the backend

Go to the main window's Settings tab:

Settings tab

  1. Click Choose backend package folder… and pick the folder the installer set up (it contains python and backend). If the game was closed while installing, the installer already filled it in.
  2. "Paths OK" and "Up to date" mean you're set.
  3. Click the power button at the top right to start the backend. Tick "Start the backend when the plugin loads" so you don't have to every time.
  4. Leave the device on Auto: it uses the GPU when there is enough VRAM and falls back to the CPU otherwise.

The status icons from left to right are Brio, GPose, IVCS (appears once a character is selected in GPose), backend, VRAM and RAM. All green means you're ready. Hover over an icon for details.

3. The main window

Main window

The toolbar has three rows: status, the backend switch and the heart at the top right (community & support, see section 22); tools (library, save to library, extra views, character list, lighting, composition and screenshots); and the character to pose, the person in the image and the race skeleton. The race is detected from the character's appearance, so you rarely need to change it.

4. Recognize your first image and apply it

  1. Enter GPose and select the character to pose (or pick one from the dropdown in the third row).
  2. Drop an image on the Image tab, or click the dashed area to pick one from the library. The first recognition loads the model (about 10–20 s); after that it takes a few seconds per image. Recognized images are cached and reopen instantly.
  3. Check the overlay: green is the skeleton recognized from the photo, orange is the FF14 skeleton for your race. If they roughly line up, you're good.
  4. Click Apply (▶) at the bottom left and the pose is on your character.

Recognition result

The result (this one also used "Match camera" below, so the camera matches the photo too):

After applying

Tips:

  • For illustrations choose Tuning → Preset → Illustration; it tones down exaggerated head and neck proportions.
  • Images dropped on the main window are not added to the library automatically; click "Save to library" in the second row if you want to keep them.

5. 3D tab: check and fine-tune joints

3D tab

  • Drag to rotate, scroll to zoom, drag with the right or middle button to pan, double-click to reset the view. Looking from the side reveals depth problems you can't see in the photo.
  • Click a joint to select it: three rotation rings appear on it (red, green and blue, one per axis; drag a ring to rotate, like in Brio), and sliders for that joint (manual offsets) appear below.

Rotation rings

  • Yellow / red dots mark joints whose angles are outside the usual in-game range and may clip.
  • When you're happy, click Save correction at the top left: the offsets and the Tuning tab's values are stored with this image and restored the next time you open it.
  • Edit in game at the top of the tab lets you drag points on the character right in GPose, and Tutorial explains it with pictures; see section 18.

6. Tuning tab: amplitude and corrections

Tuning tab

  • Amplitude: 1 = as in the photo, below 1 calmer, above 1 stronger. Right-click a slider to reset it.
  • Hand contact compensation: keeps hands on the head, chest or hips (your race's proportions differ from the person in the photo).
  • Stance: gravity correction (straightens the body when the photo's camera is tilted), snapping to the ground, and "Flatten planted feet" (feet standing on the ground are laid flat like in game; only the foot turns, the legs stay). Turn off snapping for jumps and mid-air poses.
  • Helper bones: how strongly the skin helpers at elbows, knees, shoulders etc. follow.
  • Whole-body rotation / offset: when the whole person leans or needs raising or lowering (someone sitting on a chair), rotate around the hips or move everything; camera matching turns and moves with the person, so the framing still matches the photo.
  • New images start from the defaults; keep this image's values with Save correction at the top, which also tells you about unsaved changes.

7. Hints tab

Hints tab

Lists joints outside the usual in-game range; red is more serious than yellow. The buttons below apply the suggested amplitude. (The screenshot was made offline without a game character, which is why it also says the race could not be determined. You won't see that in game.)

8. Image library

The first button in the second toolbar row opens the library.

Library

  • Import images (multi-select), whole folders, the clipboard, or drop files on the library window. Duplicates are detected by content.
  • Add folder adds an existing image folder as a source (nothing is copied).
  • Organize with folders, tags (search #tag) and favorites; rename, move, or delete to the Recycle Bin.
  • A green check means the image has been recognized; select several and right-click → Recognize to batch them.
  • Sorting: the drop-down left of the thumbnail size sorts by name, modified (the image changed or you saved corrections or lighting for it in Posegraph, whichever is latest) or added; the button next to it switches ascending / descending.
  • Double-click opens an image in the main window; right-click has extra views, save to the Brio pose library and more.

9. Extra views: fixing front/back depth

The hardest thing to tell from a single photo is depth: is the hand in front of the body or behind it, is the leg raised forward or to the side? If you have the same pose from another angle, attach it as an extra view. The plugin fuses it automatically and only corrects depth; the pose you see in the main image stays the same.

In the library, right-click the main image → Extra views (or use the extra views button in the main window's toolbar):

Extra views menu

  • Add from files / Paste from the clipboard, or attach other images selected in the library. Extra views don't show up in the library grid.
  • Images named "main-1", "main-2" are attached to "main" automatically when imported.
  • The menu shows the detected camera angle (left / right / back / almost the same angle) and warns when an image doesn't match the main pose.
  • Untick "Use extra views" on the 3D tab to compare.

What makes a good extra view:

Good Not helpful
A side view at about 90° A back view (same depth axis as the front)
Full body, the same pose Less than 30° from the main image
Camera roughly at eye level Images where an AI changed the pose (you'll get a "doesn't match" warning)

When generating a side view with AI, ask for "exactly the same pose, the same hands and feet on each side, a 90° side view with the camera at chest height", then check that left and right weren't swapped.

10. Several characters: the character list

The "character list" button in the second toolbar row opens this window, docked to the right of the main window. The top of the window has two modes: Split (one image per character) and Merge (the people in one image go to several characters).

Split: one image per character

Character list

  1. In GPose the list shows every character you can pose.
  2. Drag an image from the library onto the box to the left of a character, one image per character.
  3. Click Apply all: each image uses its own saved correction (or the defaults), and extra views are fused automatically.
  4. To fine-tune one character, click Edit: the main window switches to that character and opens its image. Save the correction on the 3D tab when you're done.

Bindings only last for the current GPose session.

Three characters, one image each, after "Apply all":

Several characters posed together

Merge: several people in one image

  1. Open the group photo in the main window. When it finds two or more people, the character list opens on "Merge" by itself; you can also click Merge at the top.
  2. Mark out each person, either way:
    • Auto draft: pick the detector next to it (Auto / Photo / Illustration) and click "Auto draft". It finds the people, recognizes them and turns the result into paint; then fix mistakes in the paint window.
    • Paint: open the paint window. Each person gets a layer; paint a rough area (the visible part plus the hidden part) and the painted area becomes that person's box. Rough is fine, but cover the whole person, including visible hands and feet.
  3. Tick the people to recognize and click Recognize (N): they are recognized together with the same image and camera.
  4. Pick a GPose actor for each person and click Apply all (N).
  5. To fine-tune one person, click Edit on their row: the main window switches to them, and each person keeps their own adjustments.
  6. Tick Replicate positions to place them as in the photo; you can set one person as the position base (their actor stays put and the others are placed around them).

If the group comes out tilted, fix it with the overall rotation / offset on the Adjust tab, or click bones to fix them. If the backend says it is too old for merge mode, update it on the settings tab and restart it.

11. Camera matching, export and restore

Buttons at the bottom of the main window:

  • Match camera (camera icon): moves the GPose camera to the photo's position, distance and angle. Needs the Posegraph HTTP service (Settings, on by default). Clicking it offers three options:
    • Normal: matches the photo as it is.
    • Roll right / roll left: for portrait shots. The game keeps rendering in landscape and the camera rolls 90° (ReShade's SSAO and depth of field don't break the way they do in portrait mode), so the photo's width spans the game's height and the person fills the screen. With roll right the person's head points left in the game, with roll left it points right; the framing guide and photo reference lie down with it, and screenshots are turned back upright. The composition window shows which mode is on and can switch back to normal.
  • Photo focal length (the button such as "25mm" next to Match camera): the full-frame equivalent focal length used for recognition; camera matching takes its angle of view and distance from it. A photo's EXIF focal length is used when present, otherwise the depth model estimates it from the image; close-ups and plain studio backdrops are often guessed too wide, and long-lens portraits aren't recognized as such. If the person looks distorted or the matched camera is too wide, pick 24 / 35 / 50 / 85 / 135mm or type one in: the photo is recognized again with it and saved in the recognition cache, so it stays next time; "Auto" switches back.
  • Restore (↶): goes back to the pose before the first apply.
  • Export (disk icon): export a .pose file, or save to the Brio pose library (with preview and tags, ready in Brio's library).

After matching, the status line shows the camera position and framing error (0.4 cm and 6.2% here):

Camera matching

12. Facial expressions (beta)

Turns the expression in the photo 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, so nothing gets distorted.

  1. The first time, type /posegraph facedump in game (or click "Export game expression data" in the facial expression section) to export the game's own expression animations once (a minute or two).
  2. On the Tuning tab open Facial expression (test), tick Apply to character, then apply.
  3. Per-part weights (recommended): eyelids, brows, mouth, jaw and gaze can be tuned separately. Adjust just the part that looks off, which works better than the overall "Strength": e.g. raise "Jaw" if the mouth doesn't open enough, lower "Gaze" if the eyes wander. Right-click a slider to reset it to 1.
  4. Head follows face: face capture corrects the head direction, so tilted and turned heads are more accurate. 0.5 by default (half each); 0 uses only the whole-body recognition, 1 follows the face fully. This works even with expressions off, and shows how far the face and the recognized head differ.

Not possible yet: sticking out the tongue (the game's animations never move it, and face capture can't measure it); a one-eyed wink and a squinting smile are sometimes confused. Face capture rarely finds faces in illustrations and then simply isn't used.

13. Keeping Brio edits in the library

If you rotate bones in Brio after applying a pose, the main window shows "Detected N bones changed in Brio" under the toolbar:

  • Keep as correction and save to library turns the edits into joint offsets on the 3D tab (the same thing as dragging bones there) and saves them in this image's correction. They come back the next time you open the image, re-applying after changing Tuning values no longer wipes them, and you can keep fine-tuning them on the 3D tab.
  • Ignore takes the current pose as the new baseline; you'll be told again only after further edits.

Poses changed with "Edit in game" (section 18) are saved into the correction as soon as you click "Done"; you don't need this bar for them.

Only body bone rotations are kept. Face bones and bone positions or scale are not; helper bones such as shoulders and elbows follow their main bones automatically, so edits to helper bones alone can't be kept either.

14. Lighting

The light bulb button in the second toolbar row. Click a setup to light the current character (loop, Rembrandt, butterfly, split, backlight and more, 11 in total). Lights are placed relative to the character and the camera. With "Follow camera" on they turn with the camera; with "Follow character" on they are placed again when Posegraph changes the character's pose or the character is moved or turned.

Lighting

  • Match the photo: estimates where the key light comes from, how hard it is, how warm or cool, and the light ratio from the shading on the face in the main window's photo, and places the lights to match. Without a face it uses the bare skin on the body, and with little bare skin it forces a rough estimate from the shading on the clothes. The first click per photo takes a few seconds.
  • Photo colors: keeps the lights where they are and only recolors them from the photo; if the photo's background isn't dark, a background light aimed behind the character is added. The swatches below show the colors picked from the photo.
  • Light area: face / half body / full body. "Auto" (the default) follows how much the game camera shows - feet visible is full body, waist visible is half body, otherwise the face. For larger areas the lights move further away and widen so the whole character is lit.
  • Light who: one person (only the character chosen in the main window); as a group (the ticked people as one group, one set of lights covering everyone, for consistent light); each person (a set of lights on each person, however spread out); auto (people within about 2.5 m share a set, people further away get their own). Tick the characters to light below; all are ticked by default and the main window's character is always lit. With many people and lights, the groups furthest from the camera are left unlit first (the window says so).
  • Adjust one set separately: with several sets, pick one under "Adjust" (at the top of the 3D lighting window and above "Adjust each light"), or click its lights in the 3D lighting view, to change only that set; "Same as the other sets" undoes the separate changes.
  • The sliders adjust overall brightness, warmth, light distance and fill (lower fill gives a higher light ratio; turn it down when the scene itself is already bright).
  • Save to this image: stores the current lights with the main window's image; they are restored whenever that image's pose is applied ("Restore with pose").
  • 3D lighting: a separate window docked to the right of the lighting window; show or hide it with "3D lighting" and drag to resize its width. It shows every light, the character and the camera: drag a light to move it, drag the diamond to turn the light head; switch between camera view, front, side and top. The ball at the top right previews the lighting ratio.
  • The buttons below the 3D view select a light; "Adjust each light" lets you solo or mute a light, change its color, intensity and angle, and add an effect (flicker, pulse, orbit, color cycle).
  • Add light: in "Adjust each light" and the 3D lighting window. Pick key, fill, rim, hair light, kicker, top, under light, catchlight, background, ambient or color accent, then adjust it; with no lights yet you can start from scratch one light at a time, and the bin next to a light's name removes it. With several people the light goes to the set chosen under "Adjust" (every set when adjusting all sets together).
  • Reset to default: above the overall sliders, and right-clicking a slider resets just that one; the button next to each light restores how it was in the setup (changed lights get a * after their name); "Reset whole setup" goes back to how the setup was when applied. Lights built from scratch have no original until you "Save as my setup"; after that, resetting goes back to the saved version.
  • Undo / redo: the lighting and 3D lighting windows share one history; Ctrl+Z / Ctrl+Y, or the two arrows left of "Re-place".
  • Status messages appear in the lighting window's title bar. Lights are removed automatically when you leave GPose.

15. Framing and screenshots

The crop button in the second toolbar row.

Composition and screenshots

  • Show framing guide draws a frame of the chosen aspect ratio over the game, dims the outside and can add guides such as the rule of thirds; switch between landscape and portrait.
  • Fit to photo takes the ratio and orientation from the photo; with the photo ratio you can still force landscape or portrait (the ratio is flipped and part of the photo is cropped). "Photo reference" overlays the photo inside the frame so you can line up the camera; you can also overlay the skeleton recognized from it.
  • In GPose it shows and sets the equivalent full-frame focal length (the 35 mm, 85 mm photographers talk about).
  • Screenshot captures the game's own render inside the frame, saved to Documents\Posegraph\Screenshots by default; the button next to it opens the folder.
  • When the camera was matched with roll right / roll left (see section 11), the framing guide and photo reference lie on their side and screenshots are turned back upright automatically; the window shows a note, and "Back to normal" only changes the guide and screenshots.

16. Video motion (beta)

Recognize the motion in a video frame by frame and play it in GPose like an animation, then correct it. The film button in the second toolbar row opens the video window.

Recognize

  1. "Open video…" (mp4 / mov / webm / mkv / avi). The first 10 seconds are selected by default: scrub to where you want and click "Set start" / "Set end".
  2. Scrub to a frame where the person is clear, click "Find people" and click the person in the frame. For couple dances, hugs or sparring you can pick several: the first one is the lead, then choose a game character for each.
  3. Pick the "Frame rate": 5 / 10 fps is enough for slow motion, use 15 / 30 for fast dancing. "Smoothing" removes frame-to-frame jitter; too much makes fast moves sluggish. Tick "Expression" to capture the face frame by frame (more conservative than for photos).
  4. "Recognize". The backend works through the frames (about half a second per frame on GPU, so a one-minute video at 10 fps takes about 5 minutes). You can close the window; the clip is saved and loaded when it's done. Clips are saved in Documents\Posegraph\Clips; open them again with "Open clip…".

Play

  • Enter GPose, choose the character and press play. "Music" plays the video's music (or an audio file with the same name in the same folder) in sync.
  • Follow: the character stays in place and the camera moves with the video's framing every frame - handheld and tracking shots look closest to the original.
  • Walk: the character moves through the scene like in the video, feet planted. "Shot" offers Original camera, Fixed camera, Smooth follow or Fixed-point pan; "When the actor is off screen" can replace the parts where the original camera looks away with an automatic camera.
  • Camera adjustments: dolly, zoom, shift, orbit, tilt and roll; the "Tracking point" sets whether the automatic camera aims at the head, chest, hips or whole body; "Keep whole body in frame" keeps raised hands and kicks in frame. Saved per clip.
  • When the playing character's race or body changes (Glamourer), the clip is rebuilt for the new skeleton automatically, without recognizing again.
  • "Save current frame to Brio pose library" saves the frame on screen as a Brio pose.

Timeline corrections

Fix what recognition got wrong (fingers especially) by hand; corrections are saved in the clip and kept when rebuilding for another race. Open "Timeline corrections" below the player:

  1. Pause on the frame to fix and move bones directly in Brio.
  2. The panel says "The pose at x s was changed in Brio"; click Keep as correction and that frame becomes a keyframe (a diamond on the track). "Revert" discards the change.
  3. Select a keyframe to change its "Range": Ease in and out (over the transition time on each side: fully corrected close to the keyframe, then blending back to the original motion) or Hold to next (held until the next keyframe of these bones). You can also remove the correction on a single bone.
  4. Copy correction copies the selected keyframe's change, which follows the original motion wherever you paste it; Copy pose copies how a bone group looks on the current frame (left hand, right hand, both hands, arms, head and neck, whole body) and pastes it as is - use this when you need the same hand gesture many times.
  5. Paste to current frame, or mark a range with "In" and "Out" and Paste to range: held across the range, fading out after the out point.
  6. On the track, click empty space to jump to that frame, drag a diamond to move it, right-click a diamond to copy or delete it; undo (Ctrl+Z) if you change your mind, redo (Ctrl+Y) if you undid too much.

Each change only rebuilds the affected frames (a few dozen frames take a second or two). The first correction after opening an older clip rebuilds the whole clip once to read the character's current bone lengths. Face bones and helper bones such as shoulders and elbows can't be kept (helper bones follow the main bones automatically).

Current limits: fingers come straight from the recognition model and are often wrong for small or blurry hands - fix them with timeline corrections; for videos with cuts, select one continuous shot; when the person is small in the frame (wide shots), motion and expressions get worse.

17. Convert to idle

Turns the pose of a library image into an idle that loops while your character stands, sits or dozes, saved as a Penumbra mod (Penumbra required). It uses the image's saved correction, so fine-tune it on the 3D / Tuning tabs first.

Convert to idle

  1. Right-click the image in the library and choose Convert to idle….
  2. Pick the Type (standing / chair sit / ground sit / doze) and the Slot: "default" is the one used without /cpose, the numbers are the /cpose variations.
  3. Tick the races to build for (male and female separately); your current character's race has a *. Some races borrow another race's animation for a slot, and the window tells you which race changes too.
  4. Click Preview position: the side / front views on the right loop the motion, grey is the original animation and teal the result. Scroll to zoom, drag to pan, double-click to reset; "Preview zoom" only exaggerates the motion in the preview so the breathing is visible.
  5. Adjust the position if needed: forward / back, up / down, left / right (cm) move the body, lean, turn, tilt (degrees) rotate it around the pelvis. Sitting and lying poses put the pelvis where the original animation has it (so they sit on the chair); standing poses are put on the floor automatically.
  6. How much each part follows the original: 0% holds the pose, 100% moves like the original. By default only the torso breathes and arms, legs and pelvis stay still (the original's arm swings and weight shifts look odd on another pose and make the feet slide).
  7. Enter a Mod name and click Build and enable. A taken name makes a new "Name 2"; saving the same image again overwrites the mod you just made; tick "Add to an existing mod with the same name" to collect several slots in one mod.

Switch to that pose in game (use /cpose for standing) to see it. In Penumbra the mod has one group per race and one option per slot, so each can be turned on or off or deleted. Facial expressions aren't part of the idle; other players only see it through a sync plugin such as Mare.

18. Edit in game (pose editor)

When a recognized pose is almost right, you don't have to fix it on the small stick figure of the 3D tab: drag points on the character right in the game. Enter GPose, select the character and click Edit in game at the top of the 3D tab. A set of dots (control points) appears on the character and the editor panel opens in the main window. An illustrated tutorial opens the first time (hold its button for 5 seconds to close it); open it again any time with Tutorial.

Editor panel

Control points

  • Blue are hands and feet: drag one and the whole arm / leg follows (IK). Green are elbows and knees: which way they bend. White are pelvis, waist, chest and head: waist / chest / head turn toward the mouse, the pelvis moves the whole body. Purple are shoulders and toes.
  • Selecting a hand shows its five finger points: drag a fingertip toward the palm to curl, outward to straighten. Selecting the head shows the gaze point; drag it to change where the eyes look.
  • A dot turning yellow / red means the joint is beyond the usual range of game animations; tick "Keep within normal range" to limit it.

Controls

  • Click to select (an orange ring appears), press and drag for free mode. Hold Shift while dragging for fine control; right-click cancels the drag.
  • With a point selected press G to move or R to rotate, then move the mouse; left-click or Enter confirms, right-click or Esc cancels. You can type a number: centimeters for moving, degrees for rotating (e.g. R, Z, 9, 0 turns 90° around Z).
  • P pins the selected hand / foot (square marker; it stays put when the body moves; both feet are pinned by default), M mirrors this side to the other, "Symmetric editing" moves both sides together. Ctrl+Z / Ctrl+Y undo / redo, Esc deselects.
  • Click Done when finished: the edits are saved into this image's correction (joint rotations, the whole-body position and the gaze) and come back the next time you open it. Cancel discards all edits of this session.

Axis lock and direction guides

During G or R, press X / Y / Z to lock an axis. A direction guide is then drawn:

  • The first press is the global axis (solid line): X is the character's left / right, Y up / down, Z front / back, no matter how the bone is tilted.
  • Pressing the same key again uses the bone's own axis (dashed line): it follows the bone's orientation, e.g. twisting the wrist along the forearm. A third press unlocks.
  • A guide line runs across the screen, and next to the point the three directions of that set of axes are drawn (red X, green Y, blue Z); the locked one has an arrow and a "+" on its positive side, with a label saying which kind of axis it is.

Axis lock: global or the bone's own

Reading and typing angles

  • While rotating, a circle is drawn around the point and the filled wedge is the angle turned, labeled in degrees; while moving, a dashed line goes back to the start with the distance in cm. Next to the mouse you also see how much X / Y / Z changed.
  • With a point selected, Rotate X / Y / Z in the panel are the bone's current angles (relative to its parent bone; X first, then Y, then Z). Drag a number left / right to adjust it or Ctrl+click to type; it can be undone like R.

The editor saves into the image's correction through the backend: after a plugin update click Update backend in Settings, otherwise the gaze and the whole-body position can't be saved.

19. Hand snapping and "Place on body part" (experimental)

Touching a face, a hand on a shoulder or the hip, holding someone's arm... poses where a hand rests on a body are hard to recognize from a single photo. The editor has two ways to put the hand there:

  • Hand snapping: with it ticked, drag a hand near a body (yours or another character in GPose) and the palm settles on it while the fingers follow the surface; hand and fingers don't sink in. The small green circle marks the contact. Hold Alt while dragging to toggle it temporarily. Fit goes from "Loose" (a small gap) to "Pressed" (palm and finger pads press into the skin a little); soft areas such as the buttocks, chest, belly and thighs are pressed a bit more automatically.
  • Place on body part… (experimental, yellow exclamation mark next to the button): select a hand, click it, move the mouse over a body to highlight the part under the cursor, and click; the hand places itself there. Esc or right-click cancels.

Recommended use of "Place on body part": first pose the hand roughly near the spot with the palm facing it, then click "Place on body part". It only refines your pose, keeping your hand direction and elbow as much as possible with the elbow hanging naturally; it moves only that arm and never bends or turns the body. When the spot is out of reach, or only reachable by twisting the wrist / shoulder unnaturally, nothing moves and the status line explains why (e.g. use the other hand, move the two closer, pose the hand roughly there first). Not happy? Ctrl+Z.

While a hand rests on a surface, P pins it to the surface (diamond marker): on your own body the hand then follows when you turn the chest or bend over.

20. Updates

  • Plugin: updates automatically through the plugin repository. The changelog pops up the first time you open the main window after an update.
  • Backend: after a plugin update the Settings tab shows the backend version. When it's older, click the download button next to it: it downloads the matching backend code and adds whichever components new features need (face capture about 64 MB, depth about 400 MB), verifying each download, so you never need to reinstall the backend. The backend is stopped first and restarted afterwards; if the new version doesn't start, the old one is restored. If GitHub is unreachable or slow, fill in a download mirror prefix; if a download breaks, just click again — finished components aren't downloaded twice.
  • Downloads fail in the plugin: download the files from GitHub in a browser or get the archives from a friend and unpack them into the backend folder yourself; see Updating the backend by hand.
  • Python / PyTorch needs replacing (the plugin tells you), you changed GPUs, or the backend is broken: run Install-Posegraph.bat again and pick the same folder; it only adds what's missing and replaces what's wrong.

Changelog

21. FAQ

Recognition is slow / not enough VRAM Keep the device on Auto; with less than 3 GB of VRAM it uses the CPU automatically. CPU recognition takes about 10–20 s per image, and recognized images are cached. Setting "Unload model" to "Keep loaded" saves the loading time but keeps the VRAM or RAM in use. "Fast recognition" is on by default (about 0.6 s for one person). "Depth-assisted recognition" makes the focal length and floor more accurate but adds about 0.5 s per image and about 1.3 GB of peak VRAM; turn it off if your machine is tight.

The backend closes right after starting, and the log says "Cannot create thread" or "System error #1455" Windows has run out of memory including virtual memory (the game alone takes over 10 GB). Enlarge the page file or set it to system-managed, or restart the game and close programs you don't need before recognizing.

Hands go into the body / don't reach the head "Hand contact compensation" on the Tuning tab keeps contact; try a larger "Full effect" distance. Otherwise drag the hand there with "Edit in game" and let hand snapping or "Place on body part" put it on (section 19), or select the joint on the 3D tab and adjust it by hand.

"Place on body part" does nothing and the status line says it can't reach / turn It moves only that arm and never bends the body or twists joints to get there. Use the other hand, or pose the hand roughly near the spot with the palm facing it first.

G / R do nothing in the editor Click a point first to select it (orange ring); while a text box of the game or another plugin has focus, keys go to that box.

A leg or arm points the wrong way front/back That's the depth ambiguity of a single photo. A side view as an extra view fixes it best; you can also adjust it on the 3D tab.

A jump gets pushed to the ground Tuning → Stance → turn off "Snap to ground".

Wrong race Pick it manually in the race dropdown at the right of the third toolbar row (happens with race-changing mods or Customize+).

Fingers are off Fingers are more accurate with the IVCS skeleton. Without it the IVCS icon is yellow and the third finger joint's bend is folded into the second.

More questions? Ask on Discord, in the QQ group or open a GitHub issue.

22. Community & support

Everything in Posegraph is free; support is entirely optional. In the plugin, click the heart at the top right of the main window or About & support at the bottom of Settings:

Support Posegraph


Sample image credits

The sample photos are CC0 (public domain) from Wikimedia Commons, downscaled:

File Original
images/sample-lean-wall.jpg Arkansas woman poses (Unsplash)
images/sample-sit-wall.jpg Smile in Air (Unsplash)
images/sample-arms-open.jpg Woman posing by drainpipe (Unsplash)
images/sample-kalaripayattu.jpg Kalaripayattu mock combat in rural Kerala