flymyai-lora-trainer

Why we included this project

LoRA training lets you teach an image model a new style or subject without retraining the whole network, and this project provides a working pipeline for Qwen-Image, Qwen-Image-Edit, and FLUX.1-dev. Configuration happens through YAML files on top of the Hugging Face diffusers stack, so teams already living in that ecosystem can get going fast. The repo also covers control-based editing for Qwen-Image-Edit, where you pair target images with control images, and includes a dataset validation script that catches missing caption files before you burn GPU hours on a bad run. A low-VRAM path targets GPUs under 24 GiB, which makes it realistic for individuals and small studios that can't rent an H200 for every experiment. The project is still in a refinement stage with test coverage pending, so treat it as a solid reference implementation to adapt rather than a turnkey product.