Developer Builds Hybrid AI and GPU Video Upscaler That Runs Entirely in the Browser

A developer working on the project tapirconvert has built a browser-based video upscaling pipeline that avoids common memory crash issues by combining two distinct processing methods. For low-resolution videos (360p–480p), the system uses an IMDN neural network via ONNX Runtime Web to reconstruct fine details that spatial algorithms cannot recover. For higher-resolution inputs like 1080p, it bypasses AI entirely and instead runs AMD's FSR 1.0 algorithm through custom WebGPU shaders, processing frames in milliseconds without overloading the GPU. The pipeline automatically routes each video to the appropriate method based on input resolution, keeping processing fully client-side without sending data to a server. The developer noted that minimizing data transfers between CPU and GPU was as critical to performance as the upscaling algorithms themselves.