First LoRA(Z-image) - dataset from scratch (Qwen2511)

This post details an individual's initial attempt at creating a LoRA (Low-Rank Adaptation) model using the Qwen-Image-Edit 2511 model. The author generated a dataset from scratch, consisting of 20 images with modest captioning, and trained the LoRA for 3000 steps. The results were surprisingly positive for a first attempt, completed in approximately 3 hours on a 3090Ti GPU. The author notes a trade-off between prompt adherence and image quality at different LoRA strengths, observing a characteristic "Qwen-ness" at higher strengths. They express optimism about refining the process and are eager to compare results between "De-distill" and Base models. The post highlights the accessibility and potential of open-source models like Qwen for creating custom LoRAs.