Spent 3 training rounds trying to get a Jean-Léon Gérôme lora to retain fini surfaces

A dedicated AI artist on Reddit shared their journey training a LoRA (Low-Rank Adaptation) to replicate the 19th-century academic painter Jean-Léon Gérôme's distinctive style. The project spanned three training rounds, each refining the captioning strategy to coax the pixelwave base model into producing Gérôme's highly finished “fini” surfaces. The first round used abstract captions focused on “structural tension,” but the model failed to connect those concepts to actual visual anchors in latent space. The second round introduced concrete descriptive anchors, improving results but revealing that pixelwave's strong brushstroke prior prevented clean, smooth surfaces. The third round doubled down on pigment-related and object-based anchors, explicitly linking materials, surfaces, edges, and light. This approach stabilized training and delivered outputs that balance Gérôme's documentary precision with theatrical staging.

The final LoRA, available on CivitAI, lets users generate images that emulate Gérôme's theatrical composition, controlled optics, and tension between stability and strain. The creator noted that this is the most time they've spent on a single training project, and they're open to discussing captioning, datasets, and artist-specific LoRA techniques with the community. The base model used was pixelwave, and the results demonstrate how careful captioning can overcome inherent biases in diffusion models to achieve a nearly impossible surface quality in AI-generated art.