How the colorized images were made

Every plate began as a high-resolution scan of an original photographic print or negative, between about 6,000 and 9,700 pixels on the long edge. That scan is the record, and it is what you see on the left of every comparison.

To produce the color version, the scan is passed to a generative model under a restoration prompt that forbids adding, removing or reinterpreting anything, and forbids inventing detail in damaged or unclear areas. The model’s output is then discarded except for its color. The finished image takes its lightness from the original scan and only its two color channels from the model.

That is the important part. Because lightness carries all the detail in a photograph — every edge, every shadow gradient, every grain of the emulsion — the finished image is geometrically and texturally the original photograph. The model contributes color and nothing else. Across the collection, measured drift between the color layer and the master is under 0.3 percent, and most plates are exact.

A further guard suppresses color where the original plate carries little information — haze, blown sky, flat distant ground — because that is where a colorizer invents most freely. Saturation is capped at a level plausible for the period. The result is deliberately restrained.

What color cannot be, is evidence. Where the original gives no basis for a hue, the choice is an informed guess constrained by what is known about period materials: Union blue stays dark, Confederate gray and butternut stay subdued, coastal sand is pale tan rather than red clay, heavy artillery is near-black cast iron rather than bronze. Treat the color as a careful reconstruction, and the black-and-white as the document.