Genko 原稿Back to edition

L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory

World models stall because memory grows with the stream. L3G-World grows memory with the scene instead: 38.6x fewer primitives and real-time generation at 9 FPS.

The explainer · 251 words

The claim

Scene-sized memory keeps world generation real-time over long exploration.

The problem

Interactive worlds must stay consistent over long sessions.

Temporal context and 3D memory both grow with the stream.

Re-encoding RGB frames for memory blocks real-time generation.

How it works

L3G memory holds compact 3D Gaussians of the scene.

A memory updater spawns Gaussians in new regions.

It refines old ones, so memory grows with the scene.

Each Gaussian stores a VAE latent feature.

The video transformer reads memory directly in latent space.

Figures

One turn in L3G-WorldUpdater and transformer alternate.

Step 1 of 4

Update memory

Spawn Gaussians where the scene is new.

Reported speed against prior 3D memoryEach figure uses the best prior baseline.

L3G-World gain

3D primitives

Up to 38.6x fewer.

Stream memory against scene memoryWhat each design costs over time.

Grows with the stream

  • Every new frame adds storage.
  • RGB frames must be re-encoded.
  • Long sessions get slower.

What changes

Memory rendering runs 359x faster than the best prior baseline.

It uses up to 38.6x fewer 3D primitives.

The system generates worlds in real time at 9 FPS.

Limits

Each figure compares with the best prior baseline per metric.

Memory still grows as the explored scene grows.

Source · arXiv

More from this edition