KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#developer-tools #codex #typescript #ai-agents #ai-agent #cli #claude-code #open-source #deepseek-harness #dsh-plugin #python #mcp

claude-skill-image-to-veo

★ 2 Open GitHub ↗

Claude Code skill that animates a static image into a 9:16 video using Google Veo 3.1. Claude reads the image and writes the cinematic prompt directly, no separate vision-model step.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# image-to-veo

A [Claude Code](https://claude.com/claude-code) skill that turns a single still image into a vertical 9:16 video clip using Google's Veo 3.1, with Claude itself acting as the creative director.

You paste an image (Dropbox link, HTTPS URL, local path, or attached file), Claude looks at it and writes a cinematic prompt grounded in what's actually in the frame, you tweak the prompt conversationally, then it generates an 8-second MP4 and drops it in your `~/Downloads/`. Built for paid-ad workflows.

## Why this is different from a "send the image to a model" workflow

Most image-to-video tools route the image through a separate vision model to generate the prompt. With Claude Code, **Claude reads the image directly** — no second API call, no second key, and you can iterate on the prompt in plain English before any credits are spent.

```
Old:  image → Gemini Vision → prompt → Veo → video
New:  image → Claude (sees it, writes prompt, you edit it in chat) → Veo → video
```

## Prerequisites

1. **[Claude Code](https://claude.com/claude-code) installed** — works in terminal, desktop, or IDE.
2. **Google AI Studio API key with Veo 3.1 access** — get one at https://aistudio.google.com/apikey. Veo 3.1 must be enabled on the project; some accounts need to be allowlisted.
3. **Python 3.7+** (any modern Mac/Linux already has this).
4. **A budget for Veo.** Generations cost roughly **$0.75/second**, so an 8-second clip is ~$6. Set the key on a project where you've reviewed billing limits.

## Install

```bash
mkdir -p ~/.claude/skills
cd ~/.claude/skills
git clone https://github.com/CurtisHowie12/claude-skill-image-to-veo image-to-veo
chmod +x image-to-veo/scripts/run.py
```

Then add your API key to your shell:

```bash
echo 'export GEMINI_API_KEY="your-key-here"' >> ~/.zshrc
source ~/.zshrc
```

(Use `~/.bashrc` instead of `~/.zshrc` if you're on bash.)

Restart Claude Code. The skill auto-registers — no further config.

## Usage

Just talk to Claude:

```
You:  animate this image: https://www.dropbox.com/scl/fi/abc/product.png?dl=0
You:  image to veo /Users/me/Pictures/hero.jpg
You:  bring this image to life [attached]
You:  make a paid-ad clip from this image
```

What happens:

1. Claude downloads the image (auto-fixes Dropbox `dl=0` → `dl=1`) and looks at it.
2. Claude writes a Veo prompt — slow cinematic camera, ambient audio matching the scene, no voiceover, with on-screen logos/headlines staying legible for ≥3 seconds.
3. **Claude shows you the prompt and waits.** Edit it conversationally: *"slower zoom"*, *"hold the headline longer"*, *"drop the bird call"*, *"more focus on the subject"*.
4. On approval, Claude calls Veo, polls every 15 seconds (jobs typically finish in ~60s), and downloads the MP4 to `~/Downloads/<your-image-name>.mp4`. If a file with that name already exists, it suffixes `-2`, `-3`, etc. — never overwrites.
5. Claude tells you the path. On macOS, `open ~/Downloads/<name>.mp4` plays it.

### Aspect ratio

Locked to **9:16** by default (vertical, paid-ad format). To override, ask: *"animate this 16:9"* or *"make it square"*.

## How it works under the hood

```
~/.claude/skills/image-to-veo/
├── SKILL.md              # frontmatter + recipe Claude follows
└── scripts/
    └── run.py            # Veo API call + 15s polling + MP4 download
```

- `SKILL.md` defines trigger phrases and the step-by-step recipe Claude executes.
- `scripts/run.py` is a single Python file using only stdlib (`urllib`, `base64`, `json`) — no third-party deps to install.
- The script reads `GEMINI_API_KEY` from the environment. **No keys are stored in the repo.**

## Cost discipline

Veo 3.1 is not cheap and there's no preview. Things this skill does to protect you:

- **Always shows the prompt before spending credits.** You approve in chat.
- **Never overwrites prior outputs.** Re-runs auto-suffix.
- **Single Veo call per generation.** No retries on prompt-quality grounds — if you don't like the result, you decide whether to spend again.

What it does *not* do:

- Track your spend. Set a billing alert in Google Cloud / AI Studio.
- Cache or reuse prior generations. Each run is fresh.

## Limitations

- **One image in, one ~8-second clip out.** No batching, no longer clips, no image-to-image continuity (yet).
- **Veo can refuse prompts.** People, brands, and certain content categories may bounce. The script surfaces Veo's error verbatim so you can adjust.
- **9:16 is opinionated.** This skill defaults to vertical paid-ad format. Override per-call as needed.

## Contributing

Issues and PRs welcome. Common improvements that would be useful:

- Batch mode (queue of images → folder of MP4s)
- Style presets (e.g. landscaping vs CPG vs fashion) selectable via flag
- Optional auto-upload step for users running Claude Code in remote sandboxes
- Other models (Veo 2, Runway, Kling) behind a `--model` flag

## License

MIT — see [LICENSE](LICENSE).

---

Built with Claude Code. Veo 3.1 is a Google product; this repo is not affiliated with Google or Anthropic.

claude-skill-image-to-veo

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README