In this work, we consider a three-stage training procedure: Alignment training, I-GPT training, and Supervised Fine-tuning.

DREAMLLM is a versatile multimodal generalist that excels at zero-shot or in-context vision language comprehension and synthesis tasks.