On-device Video Playback Upsampling
summary
Mobile playback uses Luma-only DL upscaling plus classical filters to cut latency and compute: models undergo pruning and 8‑bit quantization and leverage hardware DSP, yielding end-to-end inference under 10 ms per frame on recent phones. VSR++ shows ~30% higher quality than Lanczos in benchmarks; Lanczos slightly outperforms bicubic. An adaptive policy picks scaler by device health and resolution and integrates with ABR to stream lower‑bitrate encodings while keeping perceived quality, saving cellular data.
AI-written summary · about
chapters (8)
- 0:00Video upsampling challenges
- 1:56Modern upsampling tech overview
- 4:26Client-side upsampling
- 6:16Deep learning on mobile
- 9:06Optimizing models for mobile
- 12:17Adaptive upsampling strategies
- 13:21Upsampling with ABR
- 16:11Summary and future
related talks
10mAdventures with VP9
11mFrom Artifacts to Art: A Data-Driven Guide to Android Video Encoding
16mSwift: Adaptive Video Streaming with Layered Neural Codecs
18mScaling AV1 End-to-End Delivery at Meta
17mDelivering better video encodes for legacy devices
17mPredict, Prune, Play: Efficient Video Playback Optimization Under Device...
41m8K Codec Low-Latency Livestreaming, and AI Streaming
23mDecoder Complexity Aware AV1 Encoding Optimization
