CoDance brings multiple characters to life in sync

CoDance is an open-source animation framework designed to animate arbitrary numbers and types of characters in a single image using driving pose sequences.

While single-person image animation has become relatively mature, existing methods consistently fail when handling multiple subjects. Most systems enforce rigid, pixel-level spatial alignment between reference images and driving poses, causing character distortions or missed motions as soon as subjects are misaligned.

Unbinding motion from fixed coordinates

CoDance solves this spatial mismatch with a two-stage Unbind-Rebind pipeline.

The Unbind module introduces a pose shift encoder that adds stochastic perturbations to poses and latent representations. This forces the underlying model to discard absolute positional bias and learn location-agnostic motion.

Next, the Rebind module reassigns those learned motions to designated characters. By combining textual semantic guidance with subject-specific spatial masks, the system ensures that movements map accurately onto target bodies without bleeding into neighboring figures.

The pipeline allows a single, potentially misaligned motion sequence to drive several characters at once across distinct layouts.

Benchmarks and open release

To standardize testing for multi-character generation, the researchers also introduced CoDanceBench, evaluating models across diverse subject counts, character styles, and spatial configurations.

Experiments show that CoDance establishes state-of-the-art performance on existing benchmarks as well as CoDanceBench, handling diverse visual styles seamlessly.

The project repository and weights are already released publicly, allowing developers and creators to run and build upon the pipeline immediately.

Why it matters

Anyone working on game cutscenes, virtual influencers, or social video remixes should check this out. Animating multiple subjects previously required painstaking manual staging or separate single-character passes stitched together in post.

Technically, replacing brittle pixel-level alignment with a decoupled representation is a far more robust engineering choice. It sets a practical baseline for richer group interactions and complex animated scenes without needing perfectly staged inputs.