An AI pipeline that brings degraded black-and-white film back to life: cleaned, upscaled, colorized, and graded to HD. Methodical, evaluation-driven, and built around a custom tool I wrote where none existed.
Archival footage is a mess in inconsistent ways: wrong or unstable frame rates, duplicate and blended frames, heavy noise, flicker, instability, low resolution, and no color. Worse, every camera and film stock degrades differently, so a setting that rescues one clip ruins another.
Restoring it well is not a one-click job. It is a sequence of specialized steps, each tuned, and in the AI stages, chosen from several competing models.
I designed the whole pipeline, ran it end to end on real films, and built a custom tool for the step that needed one. The AI-heavy stages, upscaling and colorization, are treated as decisions to be tested rather than assumed. I also run the public channel the pipeline produces.
Each clip below was chosen to stress one specific stage of the pipeline. Press play and watch the before and after.
A modern series shot in black and white on today's cameras, which makes it the perfect benchmark for colorization: the source is technically pristine, so the color has nowhere to hide and every choice the pipeline makes is visible.
The surrealist short by Luis Buñuel and Salvador Dalí. Dalí is one of the artists whose work inspired The Memory Loop in the first place: the project exists to preserve the work of artists I respect and to see it in a new light. The film's pioneering visual effects also make it a demanding proving ground for the restoration process I built.
Frames from published restorations, from documentary footage of Hiroshima after the bombs and of the Bikini Atoll atomic tests. Left is the source, right is the same frame after restoration, upscaling and colorization.
Restoring the original frame rate is done with the Duplicate Frame Remover, a tool I built and open sourced; upscaling and colorization are where competing AI models are tested per clip and selected on results.
Old footage is full of duplicate and blended frames from frame-rate conversions, and off-the-shelf removers did not fit the archival workflow. So I built my own: a local web app (Node.js and Express backend, browser UI) that wraps FFmpeg's mpdecimate filter. Self-contained, no secrets, and open source.
Probes the input with ffprobe and automatically matches the output codec, profile, and pixel format (ProRes, H.264, H.265), so quality is preserved.
An adjustable threshold slider drives a live command preview that advanced users can edit directly, while the server still controls the file paths for safety.
Converts the result to a constant frame rate, which downstream editing tools require.
Returns frame-removal statistics (how many frames, what percent, how long) and manages its own upload cache.
Because each clip degrades differently, the AI stages run as comparisons. For each film I test competing models against the same footage and choose the best, often a combination, judged on detail, artifacts and temporal stability.
Topaz (Proteus, Artemis, Iris), BasicVSR++, Real-ESRGAN, and Starlight, compared per clip for detail, sharpness, and artifacts, often combined (a Proteus pass followed by an Artemis or Iris pass).
DeOldify, HAVC, and ColorMNet, compared for color accuracy and temporal stability.
Published HD restorations on the public channel, with before and after comparisons showing the jump from damaged black-and-white source to smooth, colorized HD. Featured public-domain work includes Un Chien Andalou (Luis Buñuel, 1929).
I build generative-media pipelines and the custom tools that make them work, with evaluation behind every model choice.