The AI-restored Zapruder film

The camera-original is a 25-foot spool of Kodachrome II Type A 8mm reversal stock, exposed at a measured 18.3 frames per second through a Bell & Howell Zoomatic. Its native resolution is bounded by the emulsion grain of a single 8mm frame — roughly equivalent, in modern terms, to about 700 lines of usable luminance detail. Everything a "4K restoration" adds beyond that is inference, not recovered signal (ARRB Final Report, ch. 5).

Digital stabilization — cropping and re-registering frames to remove Zapruder's handheld shake — is the least invasive intervention. It changes no pixel values, only their position. The best-known stabilized version is the 2007 rendering by researcher John Costella, which the ARRB's Roland Zavada team cited as a useful analytical tool.

AI upscaling uses a neural network trained on modern photographs to invent plausible high-frequency detail where none exists in the source. On faces and hands this often looks convincing, but the "sharpness" is generated, not measured. Two different models produce two different faces from the same Zapruder frame.

Frame interpolation — generating intermediate frames to produce 60fps footage from an 18.3fps source — invents motion the camera never saw. In the 3.5 seconds around Z-313, interpolation smooths the head-snap into a fluid arc that hides the abrupt directional reversal at the core of the forensic debate.

Every recognized panel — Warren Commission (1964), Rockefeller Commission (1975), HSCA (1979), ARRB (1998) — worked from the camera-original or first-generation prints. None used AI-restored footage, for a good reason: the analytical value of the film is that it records what a specific lens saw at a specific rate. An inferred pixel is not evidence.

For public communication — showing a general audience what Dealey Plaza looked like on Elm Street — a stabilized, gently upscaled version is genuinely helpful. For teaching, for documentaries, for holding attention long enough to explain what the frame numbers mean, modern restorations do work the grainy original cannot.

The rule is simple: use the restoration to orient the viewer; use the original to make claims. Every argument on this site that turns on a specific frame links to the NARA scan of that frame, not to a YouTube upscale.

• Frame 313 — the head shot in the camera-original
• Was the film altered? — a different alteration question
• Where the original is today