Autochromes are one of the first successful colour photography processes and are therefore significant to the history of photography. Autochromes are incredibly fragile and light sensitive, and when damaged, this damage cannot be easily reversed. As part of PERCEIVE, we are training AI models to restore images of autochrome plates exhibiting greening defects. To do so, we simulate these defects (described as synthetic defects) since we lack reference restored images.

Synthetic generation of greening defects

To train our restoration models we do not have reference or ground-truth images of their restored version. Hence, we generate these defects synthetically utilizing image processing. We first analyzed genuine greening defects and found that they typically appear either as small, point-like stains or as broader, leakage-style areas, each exhibiting multi-ringed green halos with subtle orange borders. We then reproduce these effects synthetically and apply them on digitized autochromes. The application of the type of greening defect is done randomly selecting whether an image will receive only spot defects, only large area defects, or both. For each defect, we draw an irregular oval (to mimic natural liquid spread), overlay concentric rings with slightly different green hues (and a faint orange outer ring), and apply a light Gaussian blur to soften the edges. The result is a large, diverse set of images bearing realistic-looking green marks that we can use to effectively train our restoration networks.

Example of synthetic spotting and large area greening defects

Training of deep learning network

To teach our restoration network how to remove greening stains, we first generate hundreds of paired examples—each consisting of a “damaged” image (with our synthetic green defects) and its original, non-defected version. During training, the network sees one damaged–non-defected pair at a time and adjusts its internal parameters to make its output look as close as possible to the clean image. We use an optimization method (Adam) that gradually reduces the difference between the network’s prediction and the true clean picture, while putting extra emphasis on the stained regions so that colour shifts and ring structures are corrected faithfully. As training progresses over many rounds (epochs), the model becomes increasingly skilled at erasing green spots while preserving the details and also generating the areas where greening defects are present. Finally, we evaluate the trained network on new autochrome scans it has never seen before, confirming that it can generalize from our synthetic defects to real-world damage.

Example of greened and degreened image (spotting defects)

Example of greened and degreened regions for large area defects

We can even go a step further and “re-fill” any remaining pale or off-colour spots (degreened results of spotting defects) using an inpainting model that we’ve fine-tuned with a technique called LoRA. In practice, this means the network learns from hundreds of clean autochrome examples how colours and textures should look, then automatically blends new pixels into the repaired areas.

Degreened defect corrected using inpainting

Credits

Saptarshi Neil Sinha , Julius Kühn , Johannes Koppe (Fraunhofer IGD)

Learn more

• Sinha, Saptarshi Neil, P. Julius Kuehn, Johannes Koppe, Arjan Kuijper, and Michael Weinmann. 2025. “Neural Restoration of Greening Defects in Historical Autochrome Photographs Based on Purely Synthetic Data.” Version 2. Preprint, arXiv. https://doi.org/10.48550/ARXIV.2505.22291.