Skip to content

MCVD

One model for video generation, prediction and interpolation

Visit official site

What MCVD is

The model is described as a general approach to video generation, prediction and interpolation built on a score-based diffusion loss and a 2D convolutional U-Net.

What you can do with it

  • Generate frames for film effects
  • Support video game development
  • Produce animation sequences

Model notes

  • Gaussian noise is injected into current frames for denoising
  • Past or future frames are randomly masked during training
  • Training is described with 1-4 GPUs and can scale to more channels

Pros & cons

✓ What we like

  • A simple non-recurrent 2D convolutional architecture is described
  • Samples of arbitrary length can be produced

! What to watch out for

  • Training is described on 1-4 GPUs
  • The architecture relies on conditioning on past or future frames

FAQ

What tasks are handled?

The material lists generation, future prediction, past reconstruction and interpolation.

What architecture is used?

The material describes a non-recurrent 2D convolutional U-Net architecture.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review