GF5 — Animating 3D Characters

A short martial-arts scene between a former student and his old master. The interesting work was the technical pipeline that made the scene hold together — generating, sequencing and blending motion, routing characters through the world and rendering it as a finished film.

We first built the underlying animation system: a hierarchical skeleton where child joints inherit their parents' transforms. Forward kinematics propagates local joint rotations and translations through the skeleton to give world-space joint positions.

We implemented and compared one-hot skinning, which binds each vertex to a single dominant joint, and linear blend skinning, where several joints influence a vertex through weighted transforms for smoother deformation around joints. We also visualised skinning weights to see how each joint influenced the mesh.

A custom 15-keyframe backflip was an early stress test for the whole stack — FK propagation, large joint rotations, root motion, skinning deformation and arm / leg / spine behaviour.

For the final animation, almost all fight motion was generated with HY-Motion 1.0 rather than relying only on the provided library. We:

Generating clips was not enough — they had to be composed. We separated body motion from character root paths so the same action could be positioned, rotated and timed independently in the world, letting the two fighters meet at the same contact points and react to punches, falls and knockbacks on time.

A major extension: we replaced the preset camera with a keyframed camera system. Each key controls time, position, look-at target and field of view, interpolated to pace the scene.

Personal contribution

The original scene editor only allowed root paths in X and Y. I extended it to full X, Y and Z root positions, so we could animate falls, lifts, knockbacks and characters leaving the ground.

I also debugged an FK issue where the renderer mishandled the root joint because its parent index was stored as -1, leaving characters floating at the wrong height. I fixed the root-joint handling and normalised every root waypoint to [x, y, z] — a fix that spanned animation data, FK and the renderer.

The course system used a coarse 24-joint SMPL skeleton. For the final scene we also worked on an optional 55-joint SMPL-X path adding finger chains, jaw and eyes — useful because the martial-arts motion relied on open palms, fists, grabbing and hand gestures. I worked on retargeting between the 24-joint motion representation and the richer SMPL-X skeleton.

We generated custom avatars with UP2You and connected them to the existing animation skeleton. We also adapted the asset pipeline so imported .obj / .mtl objects could be converted into the GF5 asset format and attached to character joints.

Rather than a flat background, we projected background images onto floor, walls and sky, with perspective projection and clipping so the environment stayed aligned with the camera during movement. The engineering goal was that editor preview and final render should match.

I had earlier implemented the core FK and skinning pipeline, which made it possible to debug these later issues across the entire animation stack.