FocusDiff is a new method for improving fine-grained text-image alignment in autoregressive text-to-image models. By introducing the FocusDiff-Data dataset and a novel Pair-GRPO reinforcement learning ...