Time Machine: Reconstructing perceptually, during playback, detail lost in encoding
summary
Claims real-time perceptual reconstruction of high-frequency detail lost to lossy codecs by running a stack of GPU filters and a lightweight ML model during playback, upscaling lower-resolution streams and reintroducing grain and contrast to remove banding. Implements WebGL filters that estimate missing spectrum per-frame, reports measurable gains like VMAF +4–5 points on 720p→1080p examples and subjective improvement, and includes a working WebGL demo; performance is feasible on desktop but needs tuning for mobile and cannot recover severely truncated bitstreams.
AI-written summary · about
chapters (4)
related talks
50mDemuxed 2023 Lightning Talks
20mReconstructing proprietary video streaming algorithms
18mOn-device Video Playback Upsampling
17mMorphe: High-Fidelity Generative Video Streaming with Vision Foundation Model
21mReal-Time Video Super Resolution for Live Streaming with WebGPU
19mRethinking the Video Transcoding Pipeline: How Pre-Filtering Holds the Key to Efficiency and Quality
15mReal-time video upscaling using CNNs and WebGPU
20mRealtime Video AI with Diffusion Models