Single-Image to Animatable 3D Assets: AniGen Extends AIGC for Animation and Games

Current 3D AIGC methods often produce only static meshes, which cannot be directly used in animation, game, VR, or embodied‑intelligence pipelines because they lack skeletons and skinning weights. The conventional approach separates shape generation and rigging, leading to fragile auto‑rigging that fails on noisy AIGC meshes.

AniGen (SIGGRAPH 2026) tackles this by unifying geometry, skeleton, and skinning into a single representation called S³ Fields . The model jointly predicts all three components, removing the need for a post‑processing rigging step.

The method consists of two key modules. First, a confidence‑decayed bone field predicts bone locations while explicitly modeling uncertainty in ambiguous regions, yielding cleaner and more stable skeletons. Second, a dual skinning field decouples skinning from joint count, allowing the same network to handle diverse objects such as animals, humans, plants, and mechanical arms.

Generation proceeds in two stages of flow matching. Stage 1 creates a sparse structural scaffold (bones and coarse articulation). Stage 2 refines this scaffold into high‑resolution geometry and detailed articulation, analogous to building a skeleton before adding flesh.

Experimental results on the ArticulationXL dataset show that AniGen outperforms strong baselines (TRELLIS*+UniRig, Anymate, Puppeteer, RigAnything) in both bone‑structure prediction (measured by Gromov‑Wasserstein distance) and skinning quality (Skin KL metric). Visual comparisons illustrate more stable skeletons, higher‑quality skinning, and better animation usability.

Beyond benchmarks, AniGen generalizes to a wide range of categories—people, animals, cartoon characters, plants, and mechanical devices. In‑the‑wild examples demonstrate a swimming whale, a running dog, a poseable human, a grasping robotic arm, and plants that transition between states, all with generated skeletons ready for animation.

Specific case studies include a dog model where the generated skeleton overlay confirms true animatable structure, and a robotic arm that is produced as a fully articulated object rather than a static shape. These examples highlight that the output is not merely a visual illusion of motion but a genuine, controllable 3D asset.

From a broader perspective, AniGen signals a shift in 3D generation research: future models should aim to produce directly usable, interactive digital objects instead of static shells. By integrating geometry, rigging, and skinning, AniGen brings 3D AIGC closer to practical deployment in animation production, game development, simulation, and embodied‑intelligence applications.