Back to museum

The Impressionist Museum

A generated gallery built from an SDXL LoRA experiment.

The Model

The images come from a fine-tuned SDXL workflow trained to produce Hopper-inspired scenes: quiet interiors, cinematic city light, geometric compositions, and the particular stillness that makes the source material feel architectural.

What LoRA Does

LoRA, or Low-Rank Adaptation, lets the project steer a large diffusion model without retraining every parameter. The base model stays intact while a smaller adapter learns the visual language of the training set and can be loaded into the generation pipeline.

Serving On GPUs

The production path is designed around serverless GPU inference on Modal, with generated images moving toward Modal volume storage and Postgres metadata. The ongoing infrastructure work compares that path with Kubernetes-style GPU serving, including cold starts, snapshots, LoRA loading, and cost behavior.