You can train a cnn to find bounding boxes of text first. Then run VLM on each box.