Encoding Designers’ Aesthetic Judgment into Fast Generative Interior Design: Expert-Calibrated Reward Alignment with ARIA-Diff

Fast diffusion-based interior design generation must reconcile professional aesthetic alignment with few-step inference, yet standard distillation can erase reward gains obtained during alignment. We formulate these goals as coupled constrained optimization problems, solved through penalized surrogates whose realized constraint values we audit after training, and propose ARIA-Diff, which combines an expert-calibrated Interior Aesthetic Reward Model (IARM), reward feedback alignment, and reward-preserving latent consistency distillation. IARM learns a convex combination of spatial layout, color harmony, lighting atmosphere, and material fidelity from 23,146 pairwise judgments by 43 design-trained annotators. A latent diffusion backbone adapted on 38,271 training images is optimized with semantic-fidelity and parameter-drift constraints, then distilled from 50 sampling steps to 4 through a consistency objective augmented by a hinge reward constraint. IARM predicts held-out expert preferences more accurately than the strongest tested general reward (78.2% versus 71.6%). On 1152 held-out prompts, the four-step student obtains FID 23.71, IARM 6.86, and latency 0.31 s per image, corresponding to 98.4% absolute-score retention and 84.9% alignment-gain retention relative to the 50-step teacher. In blinded comparisons, practicing designers prefer ARIA-Diff over the strongest fine-tuned baseline in 71.4% of trials. The gains hold over five training seeds, under independent preference scorers and for a separate designer panel. These results show that a shared mathematical objective can preserve domain-specific preference alignment during aggressive diffusion acceleration.