We introduce PRESLEY, an extension of the earlier ELVIS framework. Instead of destructive block removal, it applies adaptive in‑place degradation guided by a removability mask, signals per‑block degradation strength via a bit‑packed side channel, and restores content on the client using generative backbones conditioned on transmitted visual priors rather than unconditioned in‑painting.
The pipeline tackles three sub‑problems: (1) selecting blocks to degrade based on viewer attention, (2) degrading them to reduce encoder bit‑cost, and (3) reconstructing them at the receiver.
Against its predecessor at matched bitrate, PRESLEY achieves an average $-56.4\%$ BD‑rate reduction on background quality across 13 rate ladders, multiple codecs, and dataset families. Compared with pristine baselines, it delivers up to $-29.4\%$ BD‑rate savings in the bit‑starved regime, superior background quality in 17 out of 23 sequences, while preserving foreground fidelity bit‑exactly.
To map the theoretical headroom, we employ an exact leave‑one‑superblock‑out combinatorial oracle as an additive empirical bound. Existing complexity heuristics already capture about $83.3\%$ of the bit‑cost savings, leaving roughly $5\%$ of total bitrate as remaining headroom. We also identify the dominant unaddressed factor—post‑restoration damage—spanning $4.9\sim8.4\,\text{dB}$. This damage is predictable before transmission (held‑out $\rho=+0.400$), confirming the feasibility of transmit‑time restorability modeling and outlining a roadmap for joint rate‑distortion‑restoration selection rules.
Review: PRESLEY demonstrates the practical impact of generative AI in video streaming, achieving substantial bitrate reductions for background regions while keeping foregrounds intact, and paving the way for more intelligent, adaptive bitrate strategies.