Skip to content

MimicMotion

Generate human motion videos from pose sequences

Visit official site

What MimicMotion is

The model is described as a human motion video generation model jointly developed by Tencent and Shanghai Jiao Tong University, using confidence-aware pose guidance to make generation controllable, improve timing smoothness and reduce image distortion.

What you can do with it

  • Adjust the influence of pose guidance by pose estimation confidence
  • Apply region loss amplification to reduce image distortion
  • Use progressive latent fusion for long videos with smoother timing
  • Use overlapping diffusion to generate videos of arbitrary length

Application scenarios

  • Generate a dance video with fluid human motion and expressions
  • Create interactive characters with lifelike motion in virtual reality
  • Design dynamic and responsive character motions in game development

Getting started

  • Prepare a reference image and a pose sequence as input
  • Run video generation with the model
  • Adjust the parameters of confidence-aware pose guidance as needed
  • Apply region loss amplification to refine image quality in specific regions

Pros & cons

✓ What we like

  • Pose sequences condition the generation of arbitrary-length videos
  • The material cites ablation results for confidence-aware pose guidance and hand region enhancement

! What to watch out for

  • Generation is conditioned on a reference image and a pose sequence
  • Confidence-aware guidance parameters are adjusted as needed

FAQ

What guides the generation?

Confidence-aware pose guidance driven by pose estimation confidence, with a pose sequence as condition.

How long can the output be?

The material states overlapping diffusion is used for videos of arbitrary length.

What studies are mentioned?

A user study on the TikTok dataset test set and an ablation study of pose guidance and hand region enhancement.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review