Comparative analysis of image control workflows in generative AI for architectural design: a FLUX.1-based study

Abstract Text-to-image generative models often struggle to maintain visual consistency across outputs, which limits their applicability in professional architectural rendering workflows. This study examines three FLUX.1-based image-to-image workflows for architectural aerial renderings. These approaches include generation using prompts with Florence-2 providing descriptions, fine-tuning using LoRA on datasets specific to the domain, and FLUX.1 REDUX that uses image conditioning. Analysis using measures reveals that FLUX.1 REDUX shows the most consistency and maintains architectural details in the data, while generation using prompts shows high variation and fine-tuning using LoRA shows limited improvement in results. The results indicate that image-conditioned approaches provide stronger overall reference-image fidelity than text-only prompting and LoRA-based adaptation, offering a practical basis for more reliable generative workflows in architectural design.