nodecode-gemini-media

Gemini Media: pictures with Google's Nano Banana image models and clips with Gemini Omni, from a prompt or your own reference pictures. A Nodecode rewrite of bytecakelake/hermes-gemini-media-generation by bytecakelake.

The original is a backend for another agent's image_generate and video_generate lanes. Here the same two Gemini API calls are a cell of its own: three tools a model calls, one command you type, and the media written under a folder you name.

Surface	What it does
`gemini_image`	One picture from a prompt: subject, setting, style, light, framing. With `images` it edits, keeps a character, or combines products - up to 14 reference pictures. The answer names the file, its bytes, its pixel size and the model.
`gemini_video`	One clip, 3 to 10 seconds, 16:9 or 9:16, 360p to 4k, with Gemini's native audio. `image` makes that picture the first frame; `images` makes the character in them the one that moves; up to 6 in all. The answer names the file, its bytes and its real duration, read out of the MP4.
`gemini_models`	The image and video models this key can call, read from the key's own model list, with the curated list behind it when the API cannot be reached.
`/gemini-media`	The same from your side of the keyboard: `/gemini-media a red fox in snow --landscape`; `/gemini-media --video --seconds 6 a fox crossing a snowy field`. Typed alone it asks for the prompt.

Model	What it is for
`gemini-nano-banana-2.1`	The default image model: fast, 1K/2K/4K, Google Search grounding, multi-reference.
`gemini-3-pro-image`	Nano Banana Pro: complex scenes, text rendering, the highest world knowledge.
`gemini-3.1-flash-image`, `gemini-3.1-flash-lite-image`	Nano Banana 2 and 2 Lite: the cheaper image models; the lite one renders 1K only.
`gemini-2.5-flash-image`	The legacy Nano Banana, still accepted by id.
`gemini-omni-1.1-flash`	The default video model: text, image and reference to video, 3 to 10 seconds, with audio.

An image is a response_format of type image with an aspect ratio and, when you ask for one, an image size; a model that renders one size only is sent the size it renders, and the answer says when it moved. A video is a response_format of type video with an aspect ratio, a resolution and a duration the API takes as a string ("8s"). A resolution between two supported ones rounds up, and the answer says so.

Get it

Tell the agent "add the gemini-media cell", or run nodecode add nc://gemini-media.

Set it up

You need a Gemini API key from https://aistudio.google.com/apikey. Export it, and tell the cell which variable holds it:

export GEMINI_API_KEY=...

Then, in /settings (or in the gemini-media section of config.jsonc):

"gemini-media": {
  "api_url": "https://generativelanguage.googleapis.com/v1beta",
  "api_key_env": "GEMINI_API_KEY",
  "output_dir": "~/.nodecode/gemini-media"
}

Setting	Default	What it does
`api_url`	(required)	the Gemini API root. A bare host gains `/v1beta`; plain `http` only for a loopback stand-in.
`api_key`	(unset)	the key itself, pasted into `/settings`; a secret the section keeps.
`api_key_env`	`GEMINI_API_KEY`	the variable read when `api_key` is empty. `GOOGLE_API_KEY` is read when that one is unset.
`api_key_file`	(unset)	a file holding the key, which the file's own permissions guard.
`image_model`	`gemini-nano-banana-2.1`	the image model a call gets when it names none.
`video_model`	`gemini-omni-1.1-flash`	the video model a call gets when it names none.
`output_dir`	`~/.nodecode/gemini-media`	where the pictures and clips are written.
`image_timeout`	`150`	seconds one image request may take.
`video_timeout`	`300`	seconds one render may take. The video side stalls now and then; a timeout is an error, never a silent retry that bills twice.

Without a key every call answers with the variable to set; nothing is billed for a call refused before the request - an empty prompt, a missing key, a reference picture that is not a picture.

Use it

Ask in words - "draw a red fox in snow, wide shot", "make this photo look like dusk", "animate this picture, slow push in, 6 seconds" - and the model calls one of the tools. A render is slow (seconds for a picture, tens of seconds to minutes for a clip), so tell the agent to say it started before it waits.

The cell sends the prompt, the reference pictures and the key to generativelanguage.googleapis.com (or the api_url you name) and nowhere else. It reads only the picture files you pass, writes only under output_dir, runs no shell command and no self-update, and keeps no copy of the key.