Text-to-motion AI lets you lock any joint in place

MotionSpaceFlow (MSFlow) is a lightweight text-to-motion framework that turns natural language descriptions into realistic 3D human animations. Unlike standard generators that dictate the whole sequence, it lets you lock any joint to an exact 3D coordinate at any specific frame.

That means you get natural human movement without giving up director-level control over where hands and feet actually land.

What it can do

Basic text prompts work out of the box. You can ask for a high kick with the left leg, crawling forward on hands and knees, doing jumping jacks, or performing a clean cartwheel.

The real breakthrough is zero-shot spatial control. If you need a character to kick a ball or place a hand on a table, you can pin that specific joint coordinate at the exact frame.

MSFlow satisfies that constraint during inference using projection sampling, naturally generating the rest of the body's movement around it without requiring task-specific fine-tuning.

How it skips the latent bottleneck

Most existing motion diffusion models compress movement into a low-dimensional, temporally downsampled latent space learned by an autoencoder. That reduces computation, but it blurs fine temporal details and makes direct manipulation of individual frames nearly impossible.

MSFlow discards learned encoders and decoders entirely, operating directly in continuous, full-resolution motion space through representation-aware flow matching.

Its architecture, called RA-MMDiT, adapts its attention mechanism to how motion is represented. It uses causal attention for incremental frame-to-frame features to preserve forward consistency, and bidirectional attention for global XYZ coordinates to keep full-body poses structurally coherent.

Because it generates physical XYZ coordinates directly, spatial constraints can simply be enforced mathematically at each sampling step.

Hardware, specs, and setup

The pretrained model weighs only around 1.1GB on Hugging Face, meaning it runs comfortably on standard consumer GPUs.

The developers have shared inference demos, training pipelines, evaluation scripts, and pretrained checkpoints on GitHub. Setup uses uv, and a shell script is provided to test joint-constrained animations like walks and kicks sequentially on a single card.

There is one logistical catch: the repository from LY Corporation notes this is a temporary open-source release that could be converted to read-only or private later, although the code itself is marked under CC0 1.0. Getting it running also requires downloading third-party evaluation datasets like HumanML3D.

Why it matters

For indie game developers and 3D animators, MSFlow solves the biggest frustration with generative animation: lack of precision. Pure text prompts rarely align with in-game collision boxes, but letting artists pin hands and feet makes generative motion practical for interactive environments.

Technically, it also shows that ditching latent autoencoders in favor of raw coordinate flow matching avoids compression artifacts while making controllable sampling trivial. We think direct-space generation will likely become the standard pattern for physics-grounded character animation.