Colorizing Historical Photographs: The Technical Reality
Restoring color to black and white photographs from the Spanish Civil War era and Franco's regime is a process that combines algorithmic inference with heavy manual correction. When I started working with these images back in 2019, most automated tools produced results that were barely usable for serious historical work. The skin tones came out orange, the uniforms had wrong hues, and the overall composition looked like a cheap postcard rather than a documentary record. The core challenge with Franco-era imagery is the limited reference material available for training. Most AI colorization models were trained on modern photographs where clothing, skin tones, and environmental lighting fall within predictable ranges. A Spanish uniform from 1939 doesn't appear frequently enough in training datasets to get right automatically. You end up with a beige tunic that should be olive drab, or a sky that renders as an unnatural purple because the algorithm has no good anchor for Mediterranean daylight in the 1940s. Here is the workflow I settled on after spending probably four hundred hours on a single well-documented photo set:
Start with a high-resolution scan or the best available source file. Evenness matters more than anything else at this stage. A grainy, compressed JPEG will destroy any colorization attempt before it begins. I typically work from source material in the 4000-pixel range or better. Old photos of Franco often come from Spanish state archives or family collections, and the quality varies wildly depending on who preserved them and how. Run the base image through a colorization model. I've used DeOldify for years, but recent work with Stable Diffusion inpainting pipelines gives more controllable results. The initial output will be wrong in systematic ways. That is expected. The model is guessing from patterns it has seen before, and Spanish military uniforms from this period do not match those patterns well. This is where the manual work starts. I build a color reference sheet before touching the actual image. Look up the actual uniform specifications for the Spanish Army circa 1940. The field gray, the dark blue of the Navy, the brownish tones of the Moroccan Regulares — these are documented. I also cross-reference with surviving color photographs from the era, which are rare but exist. The German and Italian photographers who covered the Spanish Civil War sometimes used color film, and those images provide the ground truth you need.
I mask the image into logical regions: sky, ground, uniforms, skin, equipment, vehicles. Then I paint in colors using references, not guesses. The trick is to match the saturation and brightness of your reference colors to the local lighting conditions in the photograph. A shirt in shadow is not simply a desaturated version of a shirt in sunlight — the color temperature shifts too. I spend more time on those subtle variations than on getting the base hue right. The biggest problem I hit personally involved a well-known photo of Franco inspecting troops in 1939. The AI colorization made the entire background sky a warm golden tone, probably because it had seen too many California sunset photos in training. The actual lighting in Madrid in late spring is cooler, closer to 5600K daylight. I had to repaint the sky region entirely, then go back and adjust the color cast on the figures because they were receiving light from that sky. The uniform colors shifted slightly toward blue in the shadow areas, which the model had not captured at all. Took me about three hours to fix what the algorithm got wrong in thirty seconds. If you want to actually produce these images, here is what you need:
Get the Full Details
DeOldify is free and runs locally if you have a GPU. The repository is on GitHub and the artistic render version works well for this kind of work. For more control, the Stable Diffusion colorization pipeline with depth maps and inpainting gives better region-level accuracy but requires more technical setup. Some people use Topaz Photo AI for upscaling before colorization, which helps when working from degraded source material. The upscaling step usually cuts artifact noise significantly and makes the final output look cleaner at display size. The limitations are real. Colorization is inference, not reconstruction. You are generating plausible colors based on patterns, not recovering actual historical data. There is no way to know with certainty what a specific wall looked like in 1942 Seville. The best you can do is make an educated guess and label it as such. I always include a disclaimer on published work noting which areas are inferred versus which are grounded in documented reference material. Viewers tend to assume colorization is factual, and that assumption causes problems in historical discourse. Another issue is the ethical dimension that nobody talks about enough. Colorizing images of authoritarian figures creates a weird visual effect where the subject feels more present, more human, more immediate. It changes how people relate to the material. I have seen comments on forums where people express sympathy or fascination after seeing a colorized photo, comments they would not have made viewing the original monochrome version. That is not a problem with the technique, but it is a consequence you should be aware of before publishing this kind of work.
For source material, the Archivo General Militar de Ávila and the Hemeroteca Nacional in Madrid hold relevant collections. Some images are digitized and accessible online. Spanish family archives on platforms like Geni or Forebears sometimes contain personal photo collections that include Franco-era material. These home sources are often in better physical condition than institutional holdings because they were kept in albums rather than stored in government basements. The process for a typical image goes from raw scan to final output in about six to ten hours for someone with experience. A beginner will take significantly longer and produce less accurate results. I recommend starting with simpler images — landscapes, everyday civilian scenes — before attempting military or political subjects where color accuracy matters more for historical integrity.
