AniGen turns single images into fully rigged 3D models

Generative 3D tools have gotten scary good at building models, but they almost always spit out static digital statues. Turning those meshes into movable characters usually requires manual rigging, which is a notorious bottleneck where artists spend hours aligning virtual bones and weight-painting skins.

AniGen just changed that equation. Released on April 9, the open-source project can generate a fully functional, rigged 3D asset ready for animation straight from a single 2D image.

Bones built from the ground up

Most existing workflows treat modeling and rigging as two separate jobs. You generate a mesh, feed it into an auto-rigging plugin, and hope the bones don't puncture through the skin or mangle the limbs when moved.

AniGen flips the script by introducing unified S³ Fields, which represent Shape, Skeleton, and Skin within the exact same spatial domain. Instead of guessing bone positions after the fact, the model predicts geometry and skeletal structure together.

To make this work reliably, the researchers added a confidence-decaying skeleton field to handle ambiguous bone boundaries. They also developed a dual skin feature field that decouples skinning weights from fixed joint counts, allowing the same architecture to handle everything from a simple mechanical hinge to a complex animal skeleton.

Under the hood, a two-stage flow-matching pipeline first sketches out a sparse structural scaffold, followed by a second pass that fills in dense surface geometry and functional articulations.

From desk lamps to swimming sharks

In demonstrations, the model handled a wide range of organic and mechanical objects without breaking a sweat.

Feed it a photo of an ordinary desk lamp, and AniGen identifies the mechanical joints, delivering an articulated rig whose hinges bend naturally. Toss in a picture of a shark, and it creates a spine-aligned skeletal structure that can swim fluidly right out of the box. A photo of a dog yielded an anatomically sound rig capable of running without obvious mesh tearing.

Benchmark comparisons show AniGen closely tracking ground truth data across segmentation and skeleton estimation, handily outperforming sequential pipelines.

The project code is officially out in the open. If you want to run it on your own machine, you will need an NVIDIA GPU packing at least 18GB of VRAM.

Why it matters

AniGen marks an essential evolution for 3D generation: moving from static eye candy to usable, articulated assets.

For indie developers, modders, and solo animators, rigging is usually where projects hit a brick wall. Having a system that directly produces animation-ready meshes eliminates hours of tedious technical setup per asset. Beyond gaming and animation, this kind of automation is a huge deal for embodied AI research, where training robots requires thousands of functional, articulable everyday objects like drawers, tools, and appliances.

We think AniGen's biggest technical win is ditching the fragile post-processing habit in favor of co-generating bones and skin in latent space. The 18GB VRAM requirement keeps it out of reach for budget gaming rigs for now, and whether its skinning weights can survive extreme production deformations remains to be proven, but it makes the dream of text-to-animated-game feel surprisingly close.