Real-Time Video Super Resolution for Live Streaming with WebGPU
summary
Pipeline implements a WebGPU-powered VSR that runs on viewers' GPUs and hits real-time 60 FPS under 16.67 ms by combining bilinear pre-upscale, a small CNN on the luma channel, and optimized compute shaders. Frames enter via zero-copy import via WebCodecs and a RequestVideoFrameCallback bridge; a scout mode and 1.5x upscale 720p→1080p rollout with a performance monitor and racing dummy-upscaler keep playback smooth and show measurable increases in hours watched during live testing.
AI-written summary · about
chapters (8)
- 0:00Introduction to project
- 1:02Goals and performance
- 3:17System architecture
- 10:23Device filtering and integration
- 12:02Compute shader fundamentals
- 13:24Model and network design
- 17:16Optimization and results
- 19:05Future directions
related talks
24mSuper Resolution: The scaler of tomorrow, here today!
17mYuZu: Neural-Enhanced Volumetric Video Streaming
41m8K Codec Low-Latency Livestreaming, and AI Streaming
13mVideoCompositor: a Case Study of Live Video Composing
12mExploring Video Editing in the Browser
25mVideo AI in Your Browser with WebGPU
24mVideo streaming on the web: Live and recorded video streaming
20mRealtime Video AI with Diffusion Models