Metalsmith LoRA
A micro-dataset FLUX LoRA trained on photos of a bespoke physical ring to rapidly prototype new jewelry designs.
The Problem
I had designed and fabricated a unique copper and silver ring. I wanted to see what other pieces of jewelry (like a bangle) would look like in this exact style—before committing the time and materials to build them.
Prototyping physical jewelry is expensive and time-consuming. I realized that if I could train a model on the specific design language of my ring, I could rapidly generate and evaluate new form factors in that style. The constraint? Since it was a single physical object, I was limited to a micro-dataset of just 10 to 15 unique images.
The Execution & Direction
Training a model on such a small dataset requires immense precision. I had to ensure that the dataset curation and captioning were flawless so the model learned the *technique* and *style* rather than over-fitting to the specific shape of a ring.
I successfully trained a FLUX LoRA that captured the bespoke copper and silver aesthetic. The model extrapolated the design language perfectly, generating high-quality renders of bangles and other jewelry pieces using the very same technique.
What this proves
This project proves that AI isn't just for generating digital art; it is a highly practical tool for industrial design and rapid prototyping. By bridging my physical craftsmanship as a metalsmith with my technical capability in model fine-tuning, I created a system that drastically reduces the iteration time for physical product design.