The method
The training images were made for the project rather than scraped. Every cut, stone, metal and setting was rendered under the same studio conditions, deduplicated and checked, then used to fine-tune SDXL 1.0 with a LoRA adapter, so the model keeps one house style across every combination.
The inference
Each concept is produced in 30 steps at guidance 7.5, with negative prompts holding the composition to a single ring on a clean studio background. Every image is a generated design concept; none of them are photos of real pieces.
My part
I built and deployed the site: the studio where you design the ring, and the method page that explains how it was made, in English and Arabic.
Stack
- Field
- Generative AI, SDXL with a LoRA adapter
- Stack
