Unified Model Generates Coherent Fashion Outfits Sequentially

Generating stylistically coherent fashion outfits from vast item libraries presents a significant challenge due to the subtle and often implicit nature of aesthetic compatibility, combined with the enormous number of possible item combinations. Researchers have tackled this by formalizing the task as Constrained Ensemble Generation (CEG) and modeling it as a finite-horizon deterministic Markov Decision Process. To address CEG in the fashion domain, they propose the Unified Sequential Composition Model (USCM). This model is designed to simultaneously understand both the overall compatibility of an outfit as a set and the underlying intentions behind how items are composed. The USCM learns these crucial priors, which then guide the outfit generation process. Complementing the USCM, a Latent Expansion Monte Carlo Tree Search (LE-MCTS) mechanism is introduced. This mechanism is responsible for retrieving individual fashion items during the composition process, carefully balancing the immediate aesthetic synergy between items with the broader structural balance of the entire outfit. Extensive experiments on the Polyvore Outfits dataset, along with zero-shot evaluations on other datasets, confirm that this framework achieves state-of-the-art performance across various metrics, including human preference, automated aesthetic proxies, and structural validity.