Trading bandwidth for compute
Traditional streaming sends every encoded frame across the network. VIDER explores a different point in the design space: transmit a sparse set of keyframes, then spend GPU compute near the viewer to regenerate the missing temporal detail.
- Targeted bandwidth-constrained long-haul links.
- Separated transport cost from reconstruction cost.
- Measured the effect of different keyframe densities.
Model-agnostic pipeline
I led the design of a keyframe-plus-interpolation architecture whose network protocol does not depend on one generative model. Reconstruction models can evolve without redesigning how the sparse stream is transported.
- Used FFmpeg for extraction, encoding, and evaluation.
- Ran PyTorch reconstruction on CUDA-capable hardware.
- Integrated WAN 2.1 while keeping the reconstruction boundary replaceable.
Evaluation
The framework was evaluated across six content categories and compared with traditional codec behavior at matched bandwidth regimes to identify where generated interpolation is useful—and where its compute cost is not justified.
- Averaged 97.69% bandwidth savings across 28 test videos.
- Preserved a 0.83 average ViSiL similarity score in reconstructed video.
- Beat H.264 at CRF 50 on both bandwidth savings (97.69% vs. 96.34%) and ViSiL similarity (0.83 vs. 0.79).