Content & Media Creation · prime-skills/runcomfy-agent-skills

nano-banana-2

Name: nano-banana-2
Author: prime-skills/runcomfy-agent-skills

Generate images with Google Nano Banana 2 (Gemini-family flash-tier text-to-image) on RunComfy — bundled with the model's documented prompting patterns so the skill gets sharper output than naive prompting against the same model. Documents Nano Banana 2's strengths (rapid iteration, in-image typography rendering, predictable framing, optional web-grounded context), the resolution-tier pricing, the safety-tolerance dial, and when to route to Nano Banana Pro / GPT Image 2 / Flux 2 / Seedream instead. Calls `runcomfy run google/nano-banana-2/text-to-image` through the local RunComfy CLI. Triggers on "nano banana", "nano-banana-2", "nano banana 2", "google image gen", "gemini image", or any explicit ask to generate with this model.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandNetwork AccessWrite FileDual-pass Audit

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Dedicated single-model focus (non-router): The entire document centers on a single endpoint, `google/nano-banana-2/text-to-image`, thoroughly explaining how to utilize this model via "field-level schema + tiered resolution + prompt syntax."

nano-banana-2/SKILL.md

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

Unlike multi-model routing in video-edit/image-edit, nano-banana-2 is a "single-model deep dive" document; it shifts the responsibility of "which model to pick" upstream into a "When-to-pick" comparison table (against Nano Banana Pro / GPT Image 2 / Flux 2 / Seedream 5), while the invocation interface remains a single CLI command.

Capabilities implemented via parameter tables: num_images (1–4 batch ideation), four resolution tiers (0.5K/1K/2K/4K pricing tiers), seed reproducibility, 11 aspect_ratio values, safety_tolerance (1 strict – 6 permissive), enable_web_search, and limit_generations are all embedded with "defaults + enums + usage tips."

nano-banana-2/SKILL.md

| `safety_tolerance` | int | no | 4 | 1 (strict) – 6 (permissive). |

The "safety-tolerance dial" is directly exposed to the agent, allowing content filtering to be relaxed or tightened depending on the use case. The same table notes resolution as 0.5K (drafts)/2K (final)/4K (max), aligning with the cost-saving strategy described as "resolution-tier pricing" in the description.

Prompt value proposition = predictable in-image typography rendering and composition: The body text provides subject-first syntax + verbatim quoted text conventions, turning "uncontrollable text rendering" into actionable prompt discipline.

nano-banana-2/SKILL.md

**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.

This is the core value-add of this skill compared to "raw model invocation": subject-first declarative syntax (`Front-load subject; trail with directives`) + verbatim quoted text + fixed-seed iteration + avoiding conflicting styles (1–2 anchors).

Web grounding acts as an optional toggle rather than default behavior: enable_web_search=false is disabled by default; documentation restricts enabling it only when the prompt references current events/entities, explicitly stating the trade-offs (latency + cost).

nano-banana-2/SKILL.md

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

This rule also serves as a privacy/cost warning: once enabled, the prompt participates in online retrieval on the model's server side.

Explicit boundary conditions: Values outside the four resolution tiers and 11 aspect ratios result in a 422 error; image editing tasks are explicitly directed to the sibling nano-banana-edit (switching to the /edit endpoint).

nano-banana-2/SKILL.md

- Asking for resolutions outside the 4 tiers → 422.

The same Anti-patterns list includes `Non-quoted in-image text → unpredictable rendering`. The final section of the text notes, `For image edit (preserve subject + apply changes), see the sibling [nano-banana-edit](../nano-banana-edit) skill.`, creating an inter-family link between t2i and i2i.

Execution delegated to runcomfy CLI: `runcomfy run google/nano-banana-2/text-to-image --input '{...}' --output-dir`; the CLI sends a POST request to a fixed Model API URL, polls for completion, and downloads the artifacts.

nano-banana-2/SKILL.md

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`.

nano-banana-2 is the only file among the five surveyed that provides a complete API URL in How-it-works (the rest use domain-level descriptions)—offering the most explicit endpoint contract.

2Core Capabilities

01Flash-tier fast t2i: Tiered image generation strategy from 0.5K draft to 2K final

02Single-request 4-up batch ideation (num_images=4 + portrait aspect ratio)

03Predictable in-image typography rendering (verbatim quotes + position/font constraint syntax)

04Web-grounded current events/entity image generation (enable_web_search toggle)

05Seed-locked reproducible variation comparison

06Safety tolerance dial and single-turn generation limits (safety_tolerance / limit_generations)

07Cross-family model selection routing: Boundaries with Nano Banana Edit / Pro / GPT Image 2 / Flux 2 / Seedream 5

3External Dependencies

Kind	Dependency / Endpoint
cli	runcomfy（@runcomfy/cli 官方 CLI）
package	@runcomfy/cli（npm 全局安装）
api	Google Nano Banana 2 文生图端点（RunComfy Model API 托管）
network	生成结果下载白名单
network	enable_web_search 联网检索（模型服务端执行）
network	RunComfy 模型页 / 文档站（引用链接）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Graded/per-use authorization, sandbox recommended

For Administrators

Approval workflow, rollback support, full audit

Evidence & Triggers (File Paths & Excerpts)

网络端点: Declares external network access: model-api.runcomfy.net/v1/models/google/nano-banana-2/text-to-image — text-to-image request
本地工作目录: File I/O and artifact output: CLI downloads generated images to the local directory specified by --output-dir (path determined by caller)
SKILL.md: Declares invocation of host CLI interpreter or external runtime environment

Audit Notes: According to the unified grading criteria: The runtime drives a third-party npm package/CLI @runcomfy/cli, which reads user RunComfy credentials (~/.config/runcomfy/token.json or RUNCOMFY_TOKEN environment variable) on every invocation—triggering "invoking third-party npm/CLI with potential quota/credential access → orange". The execution surface (CLI, Model API, token reading) is entirely external to the repo, and documentation claims (0600 permissions, 2 GiB limit, no shell expansion) cannot be verified in this pin; applying the rule "when in doubt, assign the higher tier" results in an orange rating. Additionally, outbound network traffic is present: prompts (potentially containing brand or unreleased asset descriptions) are transmitted via model-api.runcomfy.net to third-party cloud infrastructure for generation; when enable_web_search=true, prompts also participate in server-side web retrieval (retrieval backends and data retention policies are undocumented). The skill itself consists purely of markdown with no scripts, obfuscation, or hidden endpoints, thus not triggering a red rating.

5Second-Pass Verification

[ok] 端点 ID 与完整 URL 一致 — description/标题用 `google/nano-banana-2/text-to-image`，How-it-works 给出同路径完整 URL，无拼写漂移。
[ok] safety_tolerance / limit_generations / enable_web_search 字段存在且默认值自洽 — safety 默认 4、limit_generations 默认 true、web_search 默认 false，正文纪律（sparingly、off by default）与默认值一致。
[unlocatable] enable_web_search 的检索后端与数据流 — 文档只给 'Adds web grounding (extra cost + latency)'，未指明由谁检索、是否外发提示词给第三方检索服务——仓内无法定位，标记待补。
[unlocatable] 『sample prompts (verified to produce strong results)』宣称 — 三个样例提示词的 'verified' 无仓内评测数据，无法复核验证过程。
[ok] 与 nano-banana-edit 的 sibling 互链 — 相对链接 `../nano-banana-edit` 在仓库内真实可达（nano-banana-edit/SKILL.md 存在），且 nano-banana-edit 反向链接本技能。
[discrepancy] 身份/安装链接一致性 — SKILL.md 安装与 GitHub 链接指向 agentspace-so/runcomfy-skills，被 pin 镜像为 prime-skills/runcomfy-agent-skills——来源标识不一致（发布身份漂移）。
[unlocatable] SKILL.md 对 CLI 行为的断言（token 0600、2 GiB 上限、无 shell 展开） — CLI 源码不在本仓库，仅文档自述。

6Verdict & Scenarios

Documents the cost strategy of "resolution pricing tiers + draft to final," preventing the agent from burning 4K by default

Actionable syntax for in-image text rendering (verbatim quotes + position/font declarations) directly targets the model's core strength

Most explicit endpoint contract: The only document across the repository providing a complete Model API URL, ensuring a clear call chain

Production parameters like safety dial, batch count, aspect ratio, and seed are fully exposed as field tables, giving the agent precise control

Upfront boundaries: 422 counterexamples and routing editing needs to nano-banana-edit reduce trial-and-error costs

Best for: Suitable for agent users with a RunComfy account who need to quickly generate marketing thumbnails, social media posts, or card posters with in-image text, especially in scenarios prioritizing "cost control via tiered resolutions + batch ideation + in-image typography"; also serves as a prompt engineering reference for Google Gemini-family image generation models.

Not for: Not suitable for tasks requiring editing existing images (preserving subjects while changing backgrounds)—that is the responsibility of nano-banana-edit (as explicitly acknowledged in the doc); not suitable for users seeking 2K/4K high-detail photorealistic masterpieces (should route to Seedream 5 / Nano Banana Pro); not suitable for users without a RunComfy account or who refuse to upload assets/prompts to the cloud; also not suitable for users requiring stable cross-version reproducibility (hosted model versions are not controlled within this repository).

Install Agent-Ready

① Verified Mirror 2026-09-09

Option A · Direct DownloadDownload nano-banana-2.tar.gz

sha256: c63e38a07d32fae6…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit fca19ae084.