# Files in prompts

> How a prompt points at your photos, videos and audio.

Source: https://docs.nocamclub.com/api/files-in-prompts

In a prompt, refer to files by their position in the request:

| In your prompt | Means |
|---|---|
| `img1`, `img2` | The first and second of `images` |
| `video1` | The first of `videos` |
| `audio1` | The first of `audios` |

```json
{
  "model": "seedance-2.5",
  "input": {
    "prompt": "img1 walks through the city from img2, moving like the dancer in video1, to the rhythm of audio1",
    "images": ["file_person", "file_city"],
    "videos": ["file_dance"],
    "audios": ["file_track"],
    "duration": 10
  }
}
```

We rewrite these into the wording each model is documented to understand, so the same prompt works everywhere. Your prompt stays as you wrote it on the generation.

## How a clip uses photos: `reference_mode`

For video models that have a choice:

| Mode | What it does | `img1` means |
|---|---|---|
| `ingredients` | The model takes what it needs: people, places, things. The clip doesn't have to start on the photo | The first photo |
| `frames` | The first photo is the opening frame, and the second, if the model takes one, is the closing frame | The first frame. `img2` is the last frame |
| `start` | The first photo opens the clip, and the rest are ingredients | The opening frame. `img2` is the first ingredient |

Videos and audio always go with `ingredients`.

## Good habits

- Say what each file is for, as in the example. A model that's told "the person from img1" does better than one left to work it out.
- Say what to keep and what to change. Anything you don't mention tends to carry over from the photo: background, clothes, light.