VQ-Style: Disentangling Style and Content in Motion with Residual Quantized Representations

How to read this

Category

Academic paper on motion style transfer through a quantized latent representation

Contributions

• Uses residual vector quantized variational autoencoders to learn a coarse to fine motion representation, on the premise that content lives in the coarse attributes and style in the finer expressive detail
• Organises content and style into separate codebooks by combining codebook learning with contrastive learning and a novel information leakage loss
• Introduces Quantized Code Swapping, an inference time technique that transfers style to unseen styles with no fine tuning
• Supports style transfer, style removal and motion blending from the same trained representation

Context

This comes from the same Disney Research and ETH group behind the generative motion rig and Two2Four, and continues their pattern of building an editable latent motion space rather than a single purpose generator. Where the motion rig gave artists sparse handles over a generative model, VQ-Style splits the latent itself so that style becomes something you can swap rather than retrain.

Correctness

The disentanglement claim rests on the architecture plus the leakage loss, so the evidence to look for is whether swapped codes genuinely leave content intact, which the paper demonstrates through its style transfer and style removal results. The no fine tuning claim for unseen styles is the strongest practical assertion and the one worth checking against the reported comparisons before you trust it on your own data.

Clarity

Written for a reader comfortable with vector quantized autoencoders and contrastive objectives. The core idea, that coarse codes carry content and fine codes carry style, is graspable without that background.

How to read it

First pass, read the abstract and the figure showing code swapping, since that single mechanism is the paper. Second pass, work through how the codebooks are separated, because the information leakage loss is what makes or breaks the disentanglement. Third pass, if you are considering this for a stylisation tool, go to the style removal results, which are the honest test of whether style and content really came apart.