rehan-remade/nvidia-kimodo

This project turns a written description of a movement, like a person dancing or walking, into an animated 3D character you can watch in your browser. It uses NVIDIA's Kimodo, a model that generates human motion from text, running on fal, a service that provides cloud GPU power on demand. When you type a prompt, the backend generates the motion as a skeleton with rotating joints, and streams the results live to the browser as they are produced, so you can see the pose forming instead of waiting for the whole animation to finish. The frontend can display the raw generated body shape, or it can apply that same motion to a fully rigged 3D character, such as an avatar imported from ReadyPlayerMe or Mixamo, standing in one of three procedurally built scenes: a forest, a plain studio, or an empty dark void. All of these scenes are built entirely in code, without needing downloaded art assets. The system has two parts. A Python backend runs on fal's cloud GPUs and handles the actual motion generation using Kimodo. A separate frontend, built with Next.js and a 3D graphics library called react-three-fiber, receives that data and renders it. Setting it up requires creating a fal account, deploying the backend so it has a working web address, and then configuring the frontend with that address and an API key. A notable technical detail is how the generated motion gets applied to different character models. Rather than copying raw rotation values, which can behave differently depending on how each character was originally built, the project points each bone in the direction the equivalent body part is moving. This makes it work with many different avatar rigs without needing custom adjustments for each one. Adding a new character is done by adding its file link to a configuration list in the code. The README does not mention a license.