Autochromes, the invention of Auguste and Louise Lumière was released publicly in 1907. It is considered the first commercially successful colour photography process. Autochromes feature a distinctive soft, muted colour palette and a unique pointillist aesthetic, derived from the potato starch granules which make up the autochrome’s colour filter. Today, displaying original autochromes is rare as original autochromes are extremely light sensitive. For this reason, it is common for digital or facsimile versions of autochromes to be presented in exhibitions.

Utilizing generative AI to generate autochrome images as well as synthetic defects

Our new approach utilizes advanced generative models to create images that attempt to capture the autochrome’s unique quality. This involves simulating common defects in autochromes like greening and cracking to develop machine learning methods for repair and enhance the quality of generated images. This approach can help us generate synthetic defects (for e.g. images with greening defects) and hence create a dataset of defected and non-defected autochrome image pairs. This would allow us to train an AI model that learns to remove these greening defects and hence potentially restore autochrome defects.

Synthetic defects generated using Text2Autochromes

To teach a diffusion model to “paint” in the distinctive autochrome style, we collected 420 high-resolution scans from early 20th-century glass-plate photographs—each paired with its original caption—resized them for GPU memory and slightly rephrased the prompts. We then fine-tuned FLUX.1-dev using Low-Rank Adaptation (LoRA) and throughout the training process, we monitored sample outputs and computed CLIPScore (a text–image compatibility metric) on varied prompts to detect overfitting. Finally, we evaluated our results via a blind user study to show that our model reliably reproduces the unique colour and texture of the historic autochrome photographs.

Credits

Julius Kühn , Duc Anh Nguyen , Saptarshi Neil Sinha (Fraunhofer IGD)

Learn more

• Kühn, Paul Julius, Saptarshi Neil Sinha, Duc Anh Nguyen, Robin Horst, Arjan Kuijper, and Dieter Fellner. 2025. “Text2Autochrome: Text Guided Autochrome Synthesis Using Generative Models.” In Digital Heritage, edited by Stefano Campana, Daniele Ferdani, Holger Graf, Gabriele Guidi, Zackary Hegarty, Sofia Pescarin, and Fabio Remondino. The Eurographics Association. https://doi.org/10.2312/dh.20253061.