Autoregressive Diffusion

Poster

View a PDF version of the poster here.

Overview

Long-term generation of realistic human motion is a difficult problem, with applications in human-computer interaction, computer animation, robotics, and more. Most approaches run the following problems:

• Unrealistic physical interactions (e.g. foot sliding)
• Repetitive or degenerate outputs (i.e. frozen motion)
• Limited interactive controllability

Current Findings:

• The main difficulty in the autoregressive mode is preserving the data’s distribution when sampling multiple times across multiple strides.
• Prenoising imposes an extra constraint across strides that helps make more consistent motions across strides.
• Inpaint training and guided inference help make the training and inference tasks more similar, improving performance.
• There remains more to be desired with the strength of trajectory control and real-time diffusion.

Publication

In the works.