Sulest LoRA

This personal project was a focused skill-building sprint to learn LoRA training end-to-end for Flux. Over a weekend, I trained my first LoRA, then ran multiple iteration cycles to improve results—treating it like a small production pipeline rather than a one-off experiment. The work emphasized practical fundamentals: curating and cleaning a training set, testing different training settings, and evaluating outputs against clear criteria (style adherence, subject consistency, artifact reduction, and prompt responsiveness).

I also developed a repeatable workflow for comparing checkpoints and making targeted adjustments—balancing model strength vs. flexibility, refining prompt templates, and stress-testing the LoRA across varied lighting, poses, and compositions to understand generalization. The outcome wasn’t just “a LoRA that works,” but a deeper, hands-on understanding of how to diagnose failure modes and systematically improve them—skills that translate directly into building reliable, brand-consistent generative pipelines.