The Cost of Dreaming: A Survey of Computational Constraints in Generative and Latent World Models
Year: 2026
Type: article
Abstract: Recent generative architectures such as Genie3 achieve 24 Frames Per Second at 720p resolution when served from Google Tensor Processing Unit clusters, requiring cloud deployment that introduces latency challenging real-time control in safety-critical robotics. Despite rapid architectural progress, computational efficiency metrics are inconsistently reported across the field; inference latency, parameter counts, and hardware requirements are disclosed selectively, while memory footprint (reported for only 40% of surveyed models) and power consumption (0% of surveyed models) remain largely unmeasured. This reporting gap obscures the efficiency-accuracy frontier and prevents practitioners from making informed deployment decisions. This paper surveys 15 representative world models spanning generative (pixel-space) and latent (joint-embedding) architec-tures published 2023-2025. We organize these models into a two-tier taxonomy: Tier 1 baseline paradigms (generative vs. latent) and Tier 2 optimization strategies (compression, acceleration, edge deployment). We establish an efficiency frontier spanning five orders of magnitude: from ultra-fast action selection (Deep AIF: 2.37 ms) to multi-second generative models (Cosmos: 383 s per sequence). Our analysis reveals efficiency-optimized models achieving 4.4-75× speedups through latent-space prediction, token compression, and architectural parallelization. We find that high-dimensional sequential operations inside the synchronous control loop appear to be a key bottleneck within the surveyed models, potentially limiting real-time performance. We outline three research priorities toward sub-100 milliseconds on-device models: hierarchical "fast-slow" architectures, dynamics-aware quantization, and standardized efficiency reporting. Our systematic comparison provides actionable deployment guidelines for robotics practitioners navigating the efficiency-expressiveness trade-off.
Authors: Hasaan Ahmad, Haixia Liu
Institution: University of the West of England
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204147246
