CLNA: controllable latent space of non-prompt attributes for enhancing diversity in text-to-image synthesis

Text-to-image synthesis in medicine remains challenging because clinically relevant visual attributes are often not fully specified by text prompts, which can limit the controllability and diversity of generated images. We propose Controllable Latent Space of Non-Prompt Attributes (CLNA), a framework that models non-prompt visual attributes within a continuous latent space for diagnosis-conditioned image synthesis, implemented with a style encoder and dual cross-attention within a Stable Diffusion backbone. We evaluated CLNA on otoendoscopic image synthesis for three diagnostic classes: normal tympanic membrane, otitis media with effusion, and chronic otitis media. Compared with the baseline Stable Diffusion model, CLNA improved image fidelity in the domain-adapted feature space, reducing Fréchet Inception Distance from 52.91 to 10.11 against the held-out internal test set, and improved feature-space recall from 57.67% to 90.22%. In classification experiments, replacing 10–40% of the real training data with CLNA-generated images at \:γ\:=0.5 gave F1 scores of 93.57–94.08% on the internal test set (vs. 94.46% averaged across 10 real-only runs) and 84.81–85.30% on the external test set (vs. 83.76% from a single real-only run). With complete replacement of the training data by generated images, CLNA outperformed the baseline model on both the internal test set (78.53% vs. 70.25%) and the external test set (72.37% vs. 54.26%). When the synthetic-only training set was increased 10-fold, CLNA further improved F1 scores to 83.21% on the internal test set and 76.77% on the external test set. These findings suggest that CLNA can control the magnitude of sampled latent variation through a style parameter, thereby improving diversity in otoendoscopic text-to-image synthesis. CLNA-generated images could support downstream classifier training under reduced use of real images.