nano-banana-2（纳米香蕉 2）

Nano Banana 2 — Pro Pack on RunComfy

runcomfy.com · Model page · GitHub

Google Nano Banana 2 — the flash-tier text-to-image model in the Gemini family — hosted on the RunComfy Model API. Optimized for ideation, social-thumbnail batches, and rapid drafts with strong in-image typography.

npx skills add agentspace-so/runcomfy-skills --skill nano-banana-2 -g

When to pick this model (vs siblings)

Nano Banana 2 is the flash-tier of the Google image-gen line. Pick it when iteration speed and predictable framing matter more than maximum detail.

If the user said "Nano Banana" / "nano-banana-2" / "Gemini image" explicitly, route here regardless. If they said "Nano Banana" without specifying 2 vs Pro, default to Pro for portraits and 2 for everything else.

Prerequisites

1. RunComfy CLI — npm i -g @runcomfy/cli
2. RunComfy account — runcomfy login opens a browser device-code flow.
3. CI / containers — set RUNCOMFY_TOKEN=<token> instead of runcomfy login.

Endpoints + input schema

google/nano-banana-2/text-to-image

For image edit (preserve subject + apply changes), see the sibling nano-banana-edit skill.

How to invoke

Default draft (1K, square, png):

runcomfy run google/nano-banana-2/text-to-image \
  --input '{"prompt": "<user prompt>"}' \
  --output-dir <absolute/path>

Vertical 4-up batch for ideation:

runcomfy run google/nano-banana-2/text-to-image \
  --input '{
    "prompt": "<user prompt>",
    "num_images": 4,
    "aspect_ratio": "9:16",
    "resolution": "0.5K"
  }' \
  --output-dir <absolute/path>

Final at 2K with seed lock:

runcomfy run google/nano-banana-2/text-to-image \
  --input '{
    "prompt": "<user prompt>",
    "resolution": "2K",
    "aspect_ratio": "16:9",
    "seed": 42
  }' \
  --output-dir <absolute/path>

Web-grounded (current event / real entity):

runcomfy run google/nano-banana-2/text-to-image \
  --input '{
    "prompt": "<prompt referencing a real-world event from this week>",
    "enable_web_search": true
  }' \
  --output-dir <absolute/path>

Prompting — what actually works

Subject-first declarative grammar. "A cinematic close-up portrait of an American woman standing under neon lights in rainy Tokyo, shallow depth of field, reflective wet streets, ultra-detailed, realistic skin texture" — primary subject, then action, environment, style, camera. Front-load subject; trail with directives.

Exact text quoting for in-image typography. "The label reads 'AURA' in clean bold sans-serif, centered, white on black" — quote the literal characters. Specify placement and font style. Don't say "with the brand name on it" and hope.

Consistent seeds for refinement. Lock seed when iterating a single prompt across small variants — keeps composition stable.

Web-grounding, sparingly. Turn on enable_web_search only when the prompt names current events / real entities. Adds latency + cost; off by default.

Don't conflict styles. "minimalist + ornate + retro + cyberpunk" cancels. Pick 1–2 anchors.

Anti-patterns:

• Trying to verbally describe a stable subject identity — use the edit endpoint with image refs instead.
• Asking for resolutions outside the 4 tiers → 422.
• Aspect ratios outside the 11 supported values → 422.
• Non-quoted in-image text → unpredictable rendering.

Where it shines

Sample prompts (verified to produce strong results)

Cinematic portrait (page example):

A cinematic close-up portrait of an American woman standing under neon
lights in rainy Tokyo, shallow depth of field, reflective wet streets,
ultra-detailed, realistic skin texture

Brand-asset card with quoted text:

A minimalist 16:9 product card: a matte black ceramic mug centered on a
soft warm-grey paper background, rim highlight from upper-left, the
headline "Brewed Quietly" in clean bold sans-serif top-right, balanced
negative space below, e-commerce ready, clean studio lighting

Vertical platform-native:

A 9:16 vertical hero for a wellness brand: a single ceramic teacup on a
linen runner, soft morning side-light, the words "Slow Down" in
hand-drawn serif large at the top, gentle steam rising, neutral color
palette, uncluttered

Limitations

• Still images only. No video on this endpoint.
• Max 4 outputs per request.
• Web search adds latency + cost — only enable on demand.
• 2K / 4K cost more — default to 1K unless user asked for higher.
• For image edit, use the /edit endpoint — not this one.

Exit codes

Full reference: docs.runcomfy.com/cli/troubleshooting.

How it works

The skill invokes runcomfy run google/nano-banana-2/text-to-image with a JSON body matching the schema. The CLI POSTs to https://model-api.runcomfy.net/v1/models/google/nano-banana-2/text-to-image, polls the request, fetches the result, and downloads any .runcomfy.net/.runcomfy.com URL into --output-dir. Ctrl-C cancels the remote request before exit.

Security & Privacy

• Token storage: runcomfy login writes the API token to ~/.config/runcomfy/token.json with mode 0600 (owner-only read/write). Set RUNCOMFY_TOKEN env var to bypass the file entirely in CI / containers.
• Input boundary: the user prompt is passed as a JSON string to the CLI via --input. The CLI does NOT shell-expand the prompt; it transmits the JSON body directly to the Model API over HTTPS. No shell injection surface from prompt content.
• Third-party content: image / mask / video URLs you pass are fetched by the RunComfy model server, not by the CLI on your machine. Treat external URLs as untrusted; image-based prompt injection is a known risk for any image-edit / video-edit model.
• Outbound endpoints: only model-api.runcomfy.net (request submission) and *.runcomfy.net / *.runcomfy.com (download whitelist for generated outputs). No telemetry, no callbacks.
• Generated-file size cap: the CLI aborts any single download > 2 GiB to prevent disk-fill from a malicious or runaway model output.