gemini-image

The maintained, machine-wide core for image generation and editing with Google Gemini. One zero-dependency Python file — runs on any machine with python, no per-project install. It is the canonical successor to the gemini-client plugin (strict superset; that plugin is retired in its favour).

Quick start

GEM=~/.claude/skills/gemini-image/scripts/gemini_image.py

# Generate
python "$GEM" --image -o castle.png "A watercolor painting of a medieval castle"

# Pick a model / aspect ratio / resolution
python "$GEM" --image -o hero.png --model gemini-3-pro-image \
    --aspect-ratio 2:3 --image-size 2K "Botanical oil painting of peonies"

# Edit or compose from reference image(s) — repeat -i (limits are per-model, see table below)
python "$GEM" --image -o staged.png -i room.png -i sofa.png \
    "Place this sofa in this room with warm afternoon light"

# See which image models your key can use
python "$GEM" --list-models --images-only

Auth resolves in order: --api-key → GEMINI_API_KEY → GOOGLE_API_KEY → .env (cwd, then next to the script). The key is sent as x-goog-api-key.

Models (verified against the live API, 2026-07)

Aspect ratios: 1:1 2:3 3:2 3:4 4:3 4:5 5:4 9:16 16:9 21:9; gemini-3.1-flash-image adds the extreme ratios 1:4 4:1 1:8 8:1.

Sizes are per-model, not universal — the CLI accepts the union and the API is the arbiter. Measured 2026-07: only gemini-3-pro-image and gemini-3.1-flash-image genuinely deliver 2K/4K. gemini-3.1-flash-lite-image returns a 400 for anything but 1K, and gemini-2.5-flash-image accepts 2K but silently returns a 1024×1024 image.

Default model: best available, auto-selected

Omit --model (CLI) or model= (library) and the skill resolves the best image model your key can access — currently gemini-3-pro-image — instead of a fixed id. It stays current automatically: a future GA Pro model (e.g. gemini-4-pro-image) is adopted the day your key gains access, no code change. A newer flash model never displaces a GA pro one (newer ≠ higher fidelity).

• Explicit always wins: --model <id> / model="<id>" skips resolution entirely (and the /models lookup).
• Pin a default: GEMINI_IMAGE_DEFAULT=<id> hard-pins (e.g. a cost-sensitive project pins gemini-3.1-flash-lite-image); explicit --model still overrides it.
• Cost/speed: the best model costs more and is slower than the floor (~$0.134 vs ~$0.067/1K image, ~3× latency — Google list pricing, 2026-07). CLI use prints a one-line stderr notice when a non-default model is auto-selected; library use stays silent.
• Never fails: the resolution is cached per key for 24h (a temp-file keyed by a hash of the API key — the key itself is never written). If /models is unreachable it falls back to gemini-3.1-flash-image, so a network blip can never break generation or cause a surprise bill.
• Why that floor and not the cheapest model: the floor has to be non-deprecated, honour every --image-size the CLI advertises, and stay well below pro pricing. gemini-3.1-flash-lite-image is cheaper but 1K-only, so a fallback would have turned a 2K request into a hard failure; gemini-2.5-flash-image is retiring on 2026-10-02. gemini-3.1-flash-image is the only current model that clears all three.

What it does beyond the old plugin

1. Reference-image input (-i/--input-image, repeatable) — image editing and multi-image composition, which the plugin could not do.
2. Multi-image output — when the model returns several images they are all saved (out.png, out-2.png, …) instead of overwriting one file.
3. Safety-block diagnostics — a blocked/empty result reports the blockReason / finishReason / safety ratings instead of a bare "no image".
4. Verified model menu + exponential-backoff retry (429/5xx) + a GeminiError exception contract so projects can vendor the file and build thin wrappers (CLI exit 2 = image blocked, 1 = error, 0 = ok).

Importable

generate, generate_image, resolve_best_image_model, extract_text, extract_images, extract_usage, list_models, and GeminiError are public. generate_image() returns {output_paths, output_path, model, mime_type, text, blocked, diagnostic, raw}.

Calling it from another language

Non-Python callers reuse this skill by shelling out and reading stdout JSON (--json), rather than reimplementing the API shape. stderr carries human notices; exit codes are 0 ok, 2 image blocked, 1 error.

python "$GEM" --image --json -o out.png "A watercolor castle"
# stdout: {"output_paths":[…], "output_path":"out.jpg", "model":"gemini-3-pro-image", …}

Two fields exist specifically for callers and are easy to get wrong:

• output_path — read the saved path from here, never from your own -o value. The models choose their own output format, so -o out.png routinely comes back as out.jpg.
• model — the id that actually ran, after resolution. On the default path the model is chosen here, at call time, so a caller that records its own hard-coded guess for provenance will drift silently the first time resolution picks something else.

This is the right integration for build-time scripts and CLIs. A compiled server or MCP runtime can't shell out to a CLI; those conform to the contract in SPEC.md instead of sharing this code.

Notes

• Every generated image carries an invisible SynthID watermark (Google, non-optional).
• The saved file's extension matches the model's actual output format (the image models choose it — gemini-3-pro-image returns JPEG). A .png request that comes back as JPEG is saved as .jpg with a stderr note, so a file never lies about its contents.
• The request uses generationConfig.imageConfig — the shape proven across all current production callers and accepted by the live API. The newer responseFormat.image shape is tracked as a future migration in SPEC.md, which is the canonical contract this skill and the project-specific TypeScript callers (ShopForge, ShopSmith, shopforge_v4) all conform to.
• Per-project business logic (brand denylists, print-spec validation, Sharp compositing, pack rules) stays in those projects — this core only does correct generation/editing.