gemini-media

name

gemini-media

description

Generate video (Veo) and images with Google Gemini from the command line — text-to-video, image+prompt-to-video, and text-to-image. Use when the user wants to make a video or image with Gemini/Veo, animate a photo into a video, generate a video from an image and a prompt, or create an AI image via Gemini. Triggers include "用 Gemini 生成视频", "把这张图做成视频", "Veo 生成", "image to video", "generate a video with Gemini", "Gemini 生图", "generate an image with Gemini".

Gemini media generation (video + image)

Drives the real Gemini web UI (gemini.google.com) through the webai CLI to generate video (Veo) and images, then downloads the result locally. No API key — it piggy-backs on your logged-in Chrome session via the opencli browser bridge.

Prerequisites (check first, in this order)

1. opencli daemon + extension connected: opencli doctor should report [OK] Daemon, [OK] Extension, [OK] Connectivity.
2. Logged-in Gemini tab in the opencli-driven Chrome profile, on an account that has Veo / video generation access (Gemini AI Pro/Ultra). If the account lacks video, image generation still works but video will fail.
3. webai CLI available: it lives at ~/projects/webai-cli. Run via node ~/projects/webai-cli/bin/webai.js … (or webai … if npm linked).

If a prerequisite is missing, tell the user exactly what to fix (e.g. "open gemini.google.com in Chrome and sign in") rather than retrying blindly.

Commands

Let WEBAI=node ~/projects/webai-cli/bin/webai.js (or just webai).

Text → image (synchronous, ~15–30s)

$WEBAI image gemini "<prompt>" --out ./out.png --json

Prints (or with --json, returns) the saved file path. Gemini saves PNG.

Text → video (asynchronous, ~1–3 min)

# 1. Submit — returns a job id immediately (the conversation hex)
$WEBAI video gemini submit "<prompt>"        # prints e.g. 7cfffb8a66f061de

# 2. Poll + download — blocks until the video is ready, then saves the mp4
$WEBAI video gemini status <job-id> --out ./out.mp4 --json

status blocks (polls up to ~10 min) until the video renders, then downloads. Add --once to status for a single non-blocking check (exit code 3 = still generating).

Image + prompt → video (animate a photo)

$WEBAI video gemini submit "<prompt>" --image ./photo.jpg   # returns job id
$WEBAI video gemini status <job-id> --out ./out.mp4

The reference image becomes the first frame / subject; the prompt drives motion.

Output contract

• Default: the media is downloaded to --out (file or directory; defaults to the current dir). The path is printed to stdout.
• --json: structured output. For video status it also includes the original Gemini url (a contribution.usercontent.google.com link, cookie-gated).

How to run it for the user

1. Verify prerequisites (opencli doctor).
2. For video: run submit, capture the job id, then run status … --out <path> and wait — it blocks until the mp4 is downloaded. Report the saved path.
3. For image: run the single image gemini command and report the saved path.
4. On failure, read the error: "could not enter … mode" → the Gemini tab may be on a different page or not loaded; "rejected (safety filter)" → the prompt was blocked; "timed out waiting for the reference image to upload" → re-check the image path.

Notes / gotchas

• One Gemini tab is driven at a time; don't run two webai media commands against the same WEBAI_SESSION concurrently (browser contention drops the connection).
• WEBAI_SESSION env var overrides the browser session name (default webai-gemini-media).
• Set WEBAI_DOWNLOAD_DIR if Chrome's download directory isn't ~/Downloads.
• Reverse-engineering details: ~/projects/webai-cli/docs/superpowers/recon/gemini-media-recon.md.