How it works
1. Sharpness. Each frame is scored by Laplacian variance; the sharpest becomes the reference unless you pick one (★).
2. Registration. Coarse offset by phase correlation, then Lucas–Kanade refinement on an image pyramid gives each frame's position to a small fraction of a pixel.
3. Fusion. Every source pixel is “dropped” onto an upsampled grid at its true sub-pixel position (drizzle). Because each frame samples the scene at a slightly different offset, the samples fill in detail no single frame contains. A second pass down-weights samples that disagree with the consensus.
4. Deconvolution. Richardson–Lucy iterations remove the blur left by the drop kernel and the sensor pixels.
Works best with 8–30 handheld shots of a static scene, same exposure, no digital zoom. Processing happens entirely in your browser.