GeoStream: Toward Precise Camera Controlled Streaming Video Generation

arXiv 2026 59.4 method

TLDR

GeoStream enables precise metric-scale camera control in autoregressive streaming video generation using a self-refreshing 3D cache and on-policy distillation.

Reasoning

The paper introduces a novel on-policy distillation method to align training and inference distributions, addressing autoregressive drift and geometric feedback. However, it focuses narrowly on camera control for video generation rather than a comprehensive world model, and lacks explicit connections to reinforcement learning.

Read-first score

Read-first score 59.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

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

Citation impact 18%

93.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93937255

Methodology quality 18%

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

Topical relevance 29%

65.7

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 449.

Keyword Scores

video world model

world model

generative world model

interactive world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Self-refreshing 3D cache that is periodically updated online from the model's own outputs, enabling geometric conditioning beyond the initial frame frustum.
On-policy distillation where conditioning during training is rendered from the student's own generated frames, aligning train and inference distributions and mitigating autoregressive drift and geometric feedback loop.
Explicit metric-scale camera control in autoregressive streaming video generation, addressing limitations of prior non-autoregressive methods with static 3D caches.

Methodology

GeoStream maintains a self-refreshing 3D cache: it estimates depth from the most recently generated frame, unprojects to 3D, and reprojects into the target view to produce point reprojections as geometric conditioning for subsequent synthesis. Training uses on-policy distillation where the conditioning is rendered from the student's own generated frames, ensuring the model trains against the exact error distribution encountered at inference.

Key Results

Quantitative and qualitative results show that the approach substantially improves camera controllability compared to prior methods.

Limitations

Relies on accurate depth estimation from generated frames, which may introduce errors that propagate through the cache.
Online cache updates may increase computational overhead during inference.
The abstract does not explicitly discuss performance on extremely long sequences or out-of-distribution camera trajectories beyond those tested.

Resources

Paper