Q-Instruct. TL, DR: Low-level visual instruction tuning, with dataset and fine-tuned checkpoints.