# Media AI Models

> Built-in Gemini media - Nano Banana 2 images, Veo 3.1 video, image and video understanding.

A curated catalog of Google Gemini generative media models, built in on your
FlyMyAI account: Nano Banana 2 (gemini-3.1-flash-image-preview) for images and
Veo 3.1 for video. Every method returns an `agent_file` with an `id` and a
`public_url` you can feed straight into the next step (e.g. `send_photo`,
`send_video`, or a Media Editor pipeline).

## What it can do

| Method | What it does |
| --- | --- |
| `generate_image` | Text-to-image with Nano Banana 2 → a `.png` `agent_file`. Optional `aspect_ratio` (14 values, incl. extremes like 1:4, 4:1, 1:8, 8:1), `image_size` (512, 1K, 2K, 4K), `enable_web_search` / `enable_image_search` grounding, and `thinking_level` (Low / High) for complex prompts. |
| `edit_image` | Image-to-image editing with Nano Banana 2. Pass the source `image` (HTTPS URL or base64) plus up to 13 more reference images (`image1`..`image13`); same `aspect_ratio`, `image_size`, search-grounding, and `thinking_level` options. Returns the edited image as an `agent_file`. |
| `generate_video` | Text-to-video or image-to-video with Veo 3.1 → an `.mp4` `agent_file`. Supports a starting `image`, a `last_frame` to interpolate to, up to 3 `reference_images` for visual consistency, `aspect_ratio` (16:9 / 9:16), `duration_seconds` (4, 6, or 8), `resolution` (720p / 1080p / 4k), `number_of_videos`, `generate_audio`, and `negative_prompt`. Polls up to ~6 minutes. |
| `analyze_image` | Gemini vision (VLM) analysis of one or more public image URLs. Returns a text description - item, lighting, suggested context - with an optional `prompt` to focus the analysis. |
| `analyze_video` | Gemini video understanding. Pass a public YouTube link (watch, shorts, youtu.be) - it is read by Gemini directly, nothing is downloaded - or a direct HTTPS video URL (an `agent_file` `public_url`, a CDN link; up to 512 MiB). Ask for a summary, hook, spoken claims, on-screen text, products shown, key moments with timestamps; optional clip range (`start_offset_seconds` / `end_offset_seconds`), `frames_per_second`, `media_resolution` and `response_format: json`. Default model `gemini-3.8-flash`. |

`generate_video` defaults to `veo-3.1-generate-preview` (best quality); you can
also pick `veo-3.1-fast-generate-preview` (lower latency),
`veo-3.1-lite-generate-preview` (fastest), or the Veo 3.0 / 2.0 models.

## How to get credentials

None - Media AI Models is a built-in tool. The Gemini API key lives on
FlyMyAI infrastructure, so there is nothing to create or paste. Just enable it.

## Fields to fill in FlyMyAI

None.

## Troubleshooting

- **Image input not used** - `edit_image` and `analyze_image` need direct
  public HTTPS URLs (or base64). Pre-upload via `download_link` if the source
  needs auth.
- **YouTube link rejected by `analyze_video`** - only public videos work
  (not private or unlisted); a direct video URL must be reachable without
  cookies or signed query parameters.
- **Video takes a while** - Veo 3.1 polls up to ~6 minutes. Generate the video
  first, then chain its `public_url` into later steps.
- **Silent video wanted** - set `generate_audio` to `false` on `generate_video`
  when you plan to add an ElevenLabs voiceover afterwards.