Project · VR · in progress
PoseForge
I'm PoseForge, a VR creature model-builder and motion-authoring tool. My job is to take a character all the way from a raw mesh to something rigged, posed, keyframed into clips, and grown with soft-tissue flesh — by hand, and eventually entirely in VR — so the bodies and motions get authored the way they'll actually play in-game. I'm the companion to the Model Testing Platform: we share a core addon called CanonBody, the rig-agnostic skeleton-and-motion system that also powers the MTP's animation "zoo."
I run in two modes. Pose/Motion mode is the default: a rigged model gets grabbable handles — cyan for IK goals like a wrist or foot, purple for FK joints like the spine or head, green for the pelvis — that you drag to pose, key into a timeline, and play back through the real animation engine, not a preview stand-in. Body/Model mode is the other half: load a raw, unrigged mesh, drop twenty draggable joint handles onto it by hand, and I auto-skin the mesh onto that skeleton, turning any model into a poseable CanonBody character. On top of the skeleton I grow the re-hosted SCD flesh — bone, muscle, and fat capsules along the limbs and torso that you can toggle into a soft, jiggling, gravity-responsive layer, or leave rigid for cheaper preview. I also handle rotoscoping: since Godot can't decode video natively, I extract reference frames with a bundled ffmpeg and let you scrub a movable reference panel alongside your timeline.
I exist because authoring motion and body shape by feel, with your hands, in the same space the character will actually move in, produces better results than sculpting a rig in an external tool and hoping the engine-side animation matches. And because a character-builder and an animation engine that don't already agree on the skeleton format are a constant source of bugs — CanonBody is the fix, and I'm one of its two proving grounds.
What works today
The headless validation suite is thorough and entirely green: exact IK reach, mirrored posing, FK, foot-pinning, onion-skinning, undo/redo, clip naming, saving and loading pattern files byte-compatible with the MTP's runtime, the walk/idle/clip engine preview, parametric overlays like reach and squat, flesh-on-rig with the jiggle layer sagging correctly under gravity, and the full rigging pipeline — load a mesh, drag twenty joints, auto-skin thirteen parts, keep the fitted flesh, and end up with an active, posable rig with correct facing and grounding. The Pose Pipeline Update — a staged build that fits the skeleton, attaches flesh, lets you fit the flesh to the specific mesh, and finalizes — has its first two phases done: per-mode flesh/jiggle toggles, and the staged build with a flesh-fit sub-mode.
What isn't verified yet is everything visual: I haven't been run in a headset. The math and logic are proven by the self-test suite, but how the walk cycle, the flesh, the skinning, and the jiggle actually look, how grabbing and the menu feel in VR, and whether the default OpenXR action bindings match the runtime are all untested — that's explicitly flagged as the next thing to check on hardware, not something I'm claiming works. Phase 3 of the Pose Pipeline Update — flesh driving the visible mesh directly, plus fingers and grab-any-part posing — is next and not yet built. Skinning quality (heat-diffusion blend weights instead of the current two-bone approach) and folding the MTP's animation engine fully into a shared rig are on the list after that.