Fishbone: From One 3D Asset to a Million Controllable Edits

We present Fishbone, a fully automatic mesh deformation framework that generates diverse, controllable variations of any input triangle mesh. The control structure is a tuple of ribs (cross-sectional iso-contours extracted at successive geodesic levels on the surface) and a smooth spine curve threaded through their centers, derived directly from the mesh via the heat method and a robust center-finding procedure with no manual annotation, template fitting, or per-shape rigging. On top of it we define six rib-driven and three spine-driven deformation primitives, all driven by low-dimensional parameters and propagated to mesh vertices through a Gaussian-weighted skinning scheme that preserves local detail. The same handle system is simulation-ready, real-time, and interactive: lumped mass and damping are inherited from the precomputed skinning weights, the integrator runs at interactive frame rates, and user input projects back to the reduced spine through the same matrix that lifts spine motion to the mesh. Experiments show greater deformation diversity and geometric plausibility than cage-based, skeleton-driven, and neural baselines, while supporting batch processing of large asset libraries, sim-to-real diversification in robotics, and procedural content creation.

From left to right, our geometry-driven pipeline takes a raw input mesh, runs a lightweight cleaning + watertightness-repair pass to produce a clean mesh, extracts ribs as iso-contours of a heat-method geodesic field rooted at an automatically selected face of the part's AABB, kept as closed loops on solid meshes and as open polylines on thin-shell parts, evaluates spine points per rib by maximizing a flatness-+-centroid-+-parent score, generates the spine by chaining the per-rib spine points into a branching tree, and computes per-vertex skinning weights that bind the mesh to the surrounding ribs via soft-cutoff Gaussians of the projected rib distance.

The same control structure ℱ = (ℛ, 𝒮, W) then drives three downstream stages, all sharing the cached weights: a parameterized static deformation algebra of rib- and spine-driven primitives, a reduced-coordinate dynamics solver with mesh lift, and a physics-free keyframe animation pipeline.