Fine-tuning gpt-4o on image data

How did you manage to add the images in JSONL for the finetuning?

Did you add them as base64 or image url?