Context & Motivation: In ultra-low-latency teleoperation, the receiver only has a few milliseconds to synthesize a future video frame. While recent generative AI models (Diffusion, Transformers) achieve impressive visual quality, their inference latency often violates real-time constraints compared to lightweight optical-flow/warping networks (such as SDC-Net).

Main Objectives:

1. State-of-the-Art Review: Survey recent (post-2020) real-time Video Frame Extrapolation and Prediction architectures.

2. Latency vs. Quality Profiling: Benchmark 1 or 2 modern lightweight PyTorch architectures against SDC-Net on GPU hardware, measuring exact forward-pass inference times across different resolutions and extrapolation horizons ($h = 1, \dots, 5$).

3. Simulator Integration: Wrap the most promising lightweight model into the simulator’s NeuralExtrapolator module and compare its distortion decay curves ($\Delta\text{PSNR}(h)$, $\Delta\text{SSIM}(h)$) against the current baseline.