Workflow
To create the gallery’s images, I followed these steps:
I trained low-rank adaptation (LoRA) models using Flux.1-dev, one of the best open source image generation models currently available. LoRA models enable users to fine-tune a base model toward specific styles or subjects, making it more proficient at generating particular features, such as a person's face or an artistic style. I developed two separate LoRA models: one trained on Roman frescoes, primarily from Pompeii and Herculaneum, and another on Fayum mummy portraits. I opted for two distinct models rather than one to maintain the flexibility to switch between them during image creation without cross-contamination. The Fayum model excelled at creating wood panel painting effects, which was my desired final output, but typically produced basic portrait compositions. In contrast, the Roman painting model offered more diverse compositions but generated images that resembled weathered frescoes. In some cases, though, the Fayum model alone proved sufficient even for complex compositions.
I collected reference images of paintings that showed clear similarities to existing ancient artworks. For instance, we have several statuettes or frescoes that likely reflect the pose of Aphrodite in Apelles' Venus Anadyomene.
Using ControlNet checkpoints for Flux, I generated images that maintained the same outlines as these references, serving as a foundation for further modifications. I experimented with various prompts, such as "A painting of Venus Anadyomene, Venus rising from the sea, wringing her hair, waves crashing at her feet, ancient roman painting, encaustic on wood panels, classical composition, masterpiece", and various parameters.
I created numerous variations of these base images using the Fayum or fresco models through an "image to image" process, experimenting with different prompts and settings (particularly focusing on guidance and denoise parameters).
After achieving a satisfactory base image, I used inpainting techniques to refine details of hands, faces, clothing, and props, followed by final adjustments in Photoshop.
