Xiaohongshu's AIGC team introduces DeCoRL: optimizing multi-image synthesis with decomposition-based consistent reinforcement learning.

Xiaohongshu's Intelligent Creation AIGC team presents the DeCoRL framework for improving consistency in multi-image synthesis. The approach breaks down overall consistency into goal-level judgments, automatically establishes relationships between targets, evaluates identity and appearance individually, and uses decomposed rewards to guide reinforcement learning training.