You are an expert at writing subject-driven video generation prompts. I'm providing you with:
1. {image_num} reference image(s) of the subject(s) that will appear in the video (referred to as image0, image1, image2, ... in order).
2. An original video description text.

Your task is to rewrite the original description into a new format with TWO parts concatenated together:

**Part 1 - Short instruction**: A concise sentence describing who the subject(s) from the reference image(s) are, what they look like briefly, where they are, and what key action/motion they perform. Reference the subject(s) using "image0", "image1", etc. to link them to the provided reference images.

**Part 2 - Long instruction**: A detailed "Generate a video where..." paragraph that describes:
- The subject(s) from the reference image(s) with detailed appearance (hair, clothing, accessories, expression, etc.), referencing them as "the person/man/woman from image0" etc.
- The scene/environment in detail (background, lighting, objects, atmosphere).
- The motion and actions in a step-by-step temporal sequence (at the start..., then..., after that...).
- The motion should remain natural and realistic.

Requirements:
- You MUST reference each subject using "image0", "image1", "image2", etc. to correspond to the provided reference images in order.
- The appearance description of each subject must be based on what you actually see in the reference image(s). Do NOT hallucinate details not visible in the images.
- The scene, actions, and motion should be derived from the original description text, but rewritten to be more detailed and vivid.
- The output must be entirely in English.
- Return ONLY a JSON object with one key: "rewritten_text". The value should be the full rewritten text (short instruction + long instruction concatenated as one string). No extra text.

Original description:
{original_text}
