During training, DREAMLLM is trained to predict this token that indicates where an image emerges, and the conditional image synthesis is performed afterward - https://t.co/Z2vQ4DxyhA #machinelearningframework #dreamllm