Skip to content

L4GM

Generate animated 4D objects from a single-view video

Visit official site

What L4GM is

The model is described as a 4D large reconstruction model that turns single-view video into animated objects.

What you can do with it

  • Reconstruct a 4D object from a video
  • Handle long videos and high frame rate footage
  • Upsample output to a higher frame rate for smooth motion

Model notes

  • Temporal self-attention layers support time consistency
  • Training uses multi-view rendering loss at each time step
  • The dataset covers 44K objects and 110K animations

Pros & cons

✓ What we like

  • A 4D interpolation model raises the frame rate
  • Multi-view rendering loss is used in training

! What to watch out for

  • Reconstruction starts from single-view video input
  • The model builds on the pretrained LGM backbone

FAQ

What does it output?

The material states per-frame 3D Gaussian Splatting representations.

How large is the dataset?

The material lists 44K objects, 110K animations, 48 views, 12M videos and 300M frames.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review