Image Super-Resolution in JS
summary
Browser-based image upscaling that actually works: an open source pipeline using Upscaler.js and Tensorflow.js converts published ESRGAN models (via Colab) to run client-side, producing 2, 3, or 4x upscales. First inference shows a ~900 ms cold start then ~100 ms subsequent runs; chunked inference with padded tiles removes edge artifacts and avoids warmup mismatches. Models can be quantized or pruned, and the package is available on npm.
AI-written summary · about
chapters (8)
- 0:00Upscaler.js introduction
- 2:13Super-resolution for user content
- 4:19Front-end vs back-end trade-offs
- 7:47Learning new ML tech
- 11:18Converting ML models with TF.js
- 13:02Browser image upscaling demo
- 16:58Performance optimization
- 19:43Future possibilities
related talks
24mSuper Resolution: The scaler of tomorrow, here today!
27mEmbedded Deep Learning Super Resolution on GStreamer using ONNX Inference Runtime
12mENHANCE! Upscaling Images with Neural Networks
21mReal-Time Video Super Resolution for Live Streaming with WebGPU
54mDeep Learning in JS
29mImage Processing in the Browser
32mSelf-Portraits in JavaScript
16mImage Processing in JavaScript