A diffusion model that keyframes humanoids like video edits

What it does

Kimodo is a diffusion model trained on 700 hours of motion-capture data that synthesizes 3D animations for human bodies and robots. Users steer it through text prompts and a dense set of kinematic constraints—full-body poses, end-effector positions, and 2D root paths—either via a local web demo with a timeline editor or a command-line interface. The repository also includes a benchmark suite and evaluation code for measuring how well models follow combined text and constraint instructions.

The interesting bit

Rather than treating motion as a black-box text prompt problem, Kimodo behaves more like animation software: you drop keyframes and waypoints onto tracks and let the diffusion model interpolate the in-betweens. It is also skeleton-agnostic enough to drive both the SOMA human rig and the Unitree G1 robot with the same architecture.

Key highlights

• Accepts simultaneous text prompts and spatial constraints, including full-body keyframes, hand/foot end-effector rotations, and 2D ground-plane waypoints.
• Ships with an interactive timeline demo featuring real-time 3D preview, playback controls, and export to NPZ, MuJoCo CSV, or AMASS formats.
• Provides a public benchmark built on the BONES-SEED dataset specifically to test text adherence and constraint-following accuracy.
• Offers multiple checkpoints trained on either a 700-hour commercially licensed mocap corpus or the smaller 288-hour public BONES-SEED set.
• Full GPU generation demands roughly 17 GB of VRAM, though the text encoder can be offloaded to CPU to drop usage below 3 GB at a speed cost.

Caveats

• The codebase is developed primarily for Linux; Windows is explicitly noted as untested outside of Docker.
• A breaking change in March 2026 shifted all model inputs and outputs to the SOMA 77-joint skeleton, which may fracture compatibility with earlier pipelines.
• The SMPL-X variant is released under a more restrictive NVIDIA R&D Model license, unlike the other checkpoints which use the NVIDIA Open Model license.

Verdict

Researchers and technical animators who need fine-grained control over humanoid motion will find Kimodo unusually direct; casual users without high-end NVIDIA hardware or interest in kinematic rigging should probably admire from a distance.