Skip to content

Lumiere

A space-time diffusion model for text-to-video

Visit official site

What Lumiere is

The model is presented as a text-to-video diffusion system that synthesises realistic, diverse and coherent motion, using a space-time U-Net to produce the entire temporal duration in one pass.

What you can do with it

  • Synthesise video with realistic, diverse and coherent motion
  • Generate the full time span of a clip in a single pass
  • Apply the design to content creation and video editing
  • Extend to image-to-video, inpainting and stylised generation

How the design works

  • Spatial and temporal downsampling and upsampling are deployed
  • A pretrained text-to-image diffusion model is reused
  • Full-frame-rate, low-resolution video is learnt at multiple space-time scales
  • The single-pass approach is contrasted with keyframe plus temporal super-resolution methods

Pros & cons

✓ What we like

  • Single-pass design aimed at global temporal consistency
  • Multiple editing tasks described in one model

! What to watch out for

  • The material is a paper summary, not a product release
  • The listed scenario entries are placeholders without detail

FAQ

What architecture is used?

The paper introduces a space-time U-Net that generates the full temporal duration in one pass.

Which editing tasks are named?

Image-to-video, video inpainting and stylised generation are listed.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review