SculptForge Files

Maquette to performance: scan with your phone, make the scan sculptable, give a head its facial shapes and drive them with your own face.

Make sculptable (splat capture -> sculptable, animation-ready mesh)

• Splats palette > From scan > Make sculptable: one click turns the capture into a new SubTool next to the splats (which are hidden, not deleted): a watertight, manifold surface (opacity-weighted Gaussian density on a voxel grid, built in parallel z slabs with the same result on any thread count; gaps closed; floater shells dropped; space boxed in on 5 of 6 sides filled, which seals open bottoms; largest piece kept), QRemesh quads at a Game (4,000), Film (12,000) or Sculpt (2,000 + 2 subdivision levels with the scan projected onto the top) budget, and the capture's colours as polypaint or an Auto UV texture (splats inside the subject are faded out). A hidden "<name> Scan" SubTool keeps the dense surface for ProjectAll.
• Face loops: a face is looked for on the scan from the current view (depth and colour, no ML model, no download): eye corners, nose tip, mouth corners and chin become QRemesh Guides (rings round the eyes and the orbits, the lips, the nasolabial loop), mirrored, with QRemesh symmetry. Review stops after the scan to drag the landmarks (the mirror follows) or Place by hand, then Finish.
• Crop Box with draggable face handles in the viewport; options for Detail, Min Opacity, Close Gaps, Fill holes, All pieces, Keep scan. Progress bar and Cancel; Ctrl-hover help; F3 entries.
• Commands: splat_to_mesh now runs the whole pipeline (detail, preset, target, levels, face_guides, landmarks, front, up, color, texture_size, crop, fill_holes, keep_parts, keep_scan, stage scan/finish, ...; resolution still works) and adds SubTools instead of replacing the sculpt; face_guides; splat_test_capture (a built-in test head or mug capture, no download).
• 5 million splats to Film quads with face loops in about 9 s, 0.9 GB (M1 Pro).
• Guide chapter "From scan to sculptable mesh"; UI test make_sculptable.sfui; UI tests: drag_landmark verb, hover scrolls a panel to the item as click does.

Added

• Facial shapes (Tool > Facial Shapes, face_shapes_generate): the 52 Apple ARKit blend shapes, exactly named, as sculpt layers on the active SubTool, from a built-in template head (ICT-FaceKit's ICT Face Model Light, MIT): automatic or landmark-guided fit (similarity + non-rigid ICP), deformation transfer (Sumner and Popovic 2004) so the shapes follow the sculpt's proportions, mirrored Left/Right on symmetric sculpts, made on a low subdivision level and carried up with the detail. Eye SubTools get eyeLook rotations, teeth and tongue the lower jaw's motion, brows and lashes follow the skin. Face test presets and a slider per shape; face_landmarks_detect, face_preset.
• Live face: face_weights json={...} sets ARKit-named shape weights on every SubTool, re-evaluating only the vertices they move (no undo step, partial GPU upload; about 9 ms for all 52 on a 700k-vertex head on an M1 Pro).
• Face takes: play recorded face feeds (JSON frames, face_take_play), bake them (face_take_bake) into a glTF animation of the morph target weights (GLB) and USD blend shape weight time samples; saved with the project.
• new_head: a new scene with a life-size human head (with mouth and eye sockets) and its eyes.
• Blender bridge: sculpt layers arrive as shape keys with their names and weights (a .shapes file beside each .GoZ, which now holds the base shape when the mesh has layers).
• UI tests: zoom and look steps; run expands $SCRIPT_DIR.
• Splat training from captures (Splats > Train from capture): a phone video, a folder of photos, a phone-link capture or a COLMAP / nerfstudio dataset becomes a Gaussian splat that appears in the viewport after the first checkpoint and sharpens while it trains (worker thread; orbit freely). Progress over the canvas: step, PSNR, elapsed and remaining time, Good enough - stop, Cancel. Presets Fast preview and Quality. Uses Brush 0.3.0 (Apache-2.0, WebGPU: NVIDIA, AMD, Intel, Apple), downloaded on request from its GitHub release with a pinned SHA-256, or a path / SCULPTFORGE_BRUSH; COLMAP (BSD; downloaded on request on Windows) for camera positions, skipped for captures with AR poses in capture.json; ffmpeg for video, keeping the sharpest frame of each stretch. The result is centred and levelled from the camera positions, saved next to the capture and copied to the LightBox Splats tab. Dropping a video onto the window trains it. Commands splat_train and splat_train_cancel (blocking in sculptforge-cli, Python and MCP). Guide chapter "Scan an object into a splat"; a generated CC0 test capture (tools/splat/make_test_capture.py, tests/data/splat_capture).
• Capture from phone (File > Capture from phone, or the phone button in the LightBox; guide chapter "Phone capture and AR", design in docs/phone-link.md): scan the QR code and the phone's browser becomes SculptForge's camera, with no app and no account. Send a scan video or a photo burst (with coaching; for splat training), one photo (to Image to 3D), or a splat .ply from Scaniverse, Polycam or Luma; uploads are chunked and resume after Wi-Fi drops, and land in <library>/Captures/<date_time>-<kind>/ with a capture.json. View in AR shows the active SubTool or the scene on your desk (USDZ for iPhone AR Quick Look, GLB for Android Scene Viewer; reduced to 80k triangles, polypaint and textures baked, re-exported after strokes). Face capture tracks 52 ARKit expressions and the head on the phone (MediaPipe Face Landmarker) and streams them to SculptForge at about 30 fps over a WebRTC data channel, for face_weights; Record face take saves them as JSON Lines. LAN only, a new token per session, size limits, safe file names; nothing listens while the window is closed.
• Build: WebRTC through libdatachannel (MPL-2.0) and Mbed TLS (Apache-2.0), fetched and linked statically; -DSCULPTFORGE_WEBRTC=OFF builds without the live face feed.
• Website: the face capture page (website/link/), which needs HTTPS for the phone's camera.
• Services: services/link-worker, a Cloudflare Worker (free plan) that relays the face page's one encrypted connection message to the desktop at link.sculptforge.app; ntfy.sh stays as the fallback. Not deployed yet.
• "Fixes & suggestions" in the LightBox header (shown at startup) and the Help menu: opens the bug report and feature request forms on GitHub.

Changed

• The website moved to https://sculptforge.app (old islandmed.github.io/SculptForge links redirect). The app, update check, user guide, packages and Blender add-on link there.
• Website: the tagline now reads "the free, open-source alternative to ZBrush and Mudbox".

Downloads

New here? install/ explains which file to pick, or install with one command: curl -fsSL https://raw.githubusercontent.com/islandmed/SculptForge/main/install/install.sh | sh (macOS / Linux), irm https://raw.githubusercontent.com/islandmed/SculptForge/main/install/install.ps1 | iex (Windows PowerShell). Each file is also attached without its version number (e.g. SculptForge-macos.dmg) so links to releases/latest/download/... always get the newest build.

The builds are not code-signed yet. macOS: right-click the app → Open the first time. Windows: if SmartScreen appears, choose More info → Run anyway. Checksums are in SHA256SUMS.txt.