Terra: Explorable Native 3D World Model with Point Latents

arXiv 25.10 2025 49.6 method

TLDR

Terra proposes a native 3D world model using point latents for explorable environment generation with multi-view consistency.

Reasoning

Strengths include a novel point-to-Gaussian VAE and sparse point flow matching achieving state-of-the-art reconstruction and generation with 3D consistency. Weaknesses are limited to static indoor scenes, no dynamics or interaction, and evaluation only on ScanNet.

Read-first score

Read-first score 49.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 25.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,experiment

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset,github

Topical relevance 42%

35.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 420.

Keyword Scores

world model

generative world model

world simulator

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Proposes a native 3D world model (Terra) that represents and generates explorable environments in an intrinsic 3D latent space, moving beyond pixel-aligned representations.
Introduces a point-to-Gaussian variational autoencoder (P2G-VAE) that encodes 3D inputs into latent point representations and decodes them as 3D Gaussian primitives for joint geometry and appearance modeling.
Develops a sparse point flow matching network (SPFlow) for generating latent point representations by simultaneously denoising positions and features.
Achieves exact multi-view consistency with native 3D representation and architecture, enabling flexible rendering from any viewpoint with a single generation process.
Enables explorable world modeling through progressive generation in the point latent space.

Methodology

Terra uses a point-to-Gaussian variational autoencoder (P2G-VAE) to encode 3D inputs into a latent point representation, which is then decoded as 3D Gaussian primitives to jointly model geometry and appearance. A sparse point flow matching network (SPFlow) generates the latent point representation by denoising both positions and features of the point latents. The model supports progressive generation in the point latent space for explorable world modeling.

Key Results

On the ScanNet v2 indoor dataset, Terra achieves state-of-the-art performance in both reconstruction and generation tasks with high 3D consistency.

Resources

Paper Project