Text-to-Motion and AI Character Animation

Text-to-motion systems generate body poses, not pixels. Each frame describes a skeleton: a root position and orientation plus rotations for each joint, often expressed on a parametric body model such as SMPL. A clip is therefore a grid of frames by features, which is why generative techniques built for images and language can be adapted to movement. The best-known approach is the Motion Diffusion Model (MDM), published by researchers at Tel Aviv University in 2022. It starts from random noise shaped like a motion clip and uses a transformer to remove that noise step by step, steered by a text embedding from CLIP. Models like this learn from datasets that pair motion capture with written descriptions, most notably HumanML3D, whose clips were drawn largely from the AMASS motion capture collection. A second family treats motion like language: T2M-GPT and MoMask compress movement into discrete tokens with VQ-VAE-style quantization, then generate those tokens the way a language model generates words. A separate branch skips text entirely. AI motion capture tools such as Move.ai, DeepMotion, Rokoko Vision and Wonder Studio (now Autodesk Flow Studio) estimate 3D poses from ordinary video. They detect 2D body keypoints in each frame, lift them into 3D, and smooth the result over time. For many projects this replaces capture suits and marker studios. In both cases the output must be retargeted: mapped onto a character rig whose bone names, proportions and rest pose differ from the source skeleton. A common misconception is that these tools deliver finished performances. Typical flaws include foot sliding, where feet skate along the ground, feet sinking through the floor, jitter, and stiff hands, because many datasets record fingers poorly. Generated clips are usually only a few seconds long, and prompts involving long action sequences or objects remain difficult. Most studios treat the output as a first pass that an animator refines in tools like Maya, Blender or MotionBuilder.