Persistent Planar Memory for Video World Models

arXiv 2026 35.9 method

TLDR

PPMem is a video world model using plane primitives as persistent 3D memory, jointly predicting RGB and planar geometry to maintain coherence across long generations.

Reasoning

The paper introduces a novel persistent planar memory mechanism that makes video world models more compact and geometrically explicit. Strong empirical validation on DL3DV and RealEstate10K supports its claims, though the abstract does not discuss limitations or computational costs.

Read-first score

Read-first score 35.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 25%

61.4

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

Methodology quality 15%

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

Reproducibility 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 334.

Keyword Scores

video world model

world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Uses oriented plane primitives as a persistent 3D memory for video world models, requiring only 7 parameters per primitive.
Introduces PPMem, augmenting a video diffusion transformer with a planar prediction head and a Depth Token Fusion (DTF) module.
Autoregressive chunks jointly generate RGB video and planar geometry, merged into a persistent plane memory that conditions later chunks via rendered depth.
Achieves compact, geometrically explicit, cross-view consistent, and adaptive 3D memory (~100K primitives, 5MB per full scene).

Methodology

PPMem augments a video diffusion transformer with a planar prediction head and Depth Token Fusion (DTF). Each autoregressive chunk jointly produces RGB video and planar geometry, which is merged into a persistent plane memory; subsequent chunks are conditioned via depth rendered from this memory. The method is evaluated on DL3DV and RealEstate10K against prior 3D-grounded world models.

Key Results

On DL3DV and RealEstate10K, PPMem outperforms prior 3D-grounded world models in video quality while jointly producing metric-scale depth, surface normals, and exportable planar geometry, and it sustains coherence across hundreds of frames under large viewpoint changes.

Resources

Paper