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

nano-banana-edit

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

Edit images with Google Nano Banana 2 (image-to-image edit endpoint) on RunComfy. Documents Nano Banana Edit's strengths (preserve subject identity, swap background, localize edits with spatial language, multi-image batch edits up to 20 inputs), the schema, and when to route to GPT Image 2 edit / Flux Kontext / Nano Banana 2 t2i instead. Calls `runcomfy run google/nano-banana-2/edit` through the local RunComfy CLI. Triggers on "nano banana edit", "edit with nano banana", "image edit nano banana", or any explicit ask to edit 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-endpoint documentation (i2i editing): Centers around `google/nano-banana-2/edit`. Core parameters include the `image_urls` array (1–20 publicly crawlable URLs, with the first image acting as the primary anchor), `number_of_images` (1–4 outputs), `seed`, `aspect_ratio`, `resolution`, `output_format`, `safety_tolerance`, `limit_generations`, and `enable_web_search`.

nano-banana-edit/SKILL.md

| `image_urls` | array | yes | — | **1–20** publicly-fetchable HTTPS URLs. |

The only structural difference from its sibling `nano-banana-2` lies in the input interface: t2i receives a prompt, whereas this endpoint receives 'prompt + up to 20 URLs', making batch consistent editing (SKU galleries / A/B variants) the signature capability of this skill.

First prompt principle = preservation-first: Explicitly prepend 'Keep [identity/pose/clothing/brand/framing] unchanged' before describing modifications. The documentation explains the underlying mechanism (the model strictly adheres only to front-loaded preservation targets).

nano-banana-edit/SKILL.md

**Preservation first, change last.** Always lead with `"Keep [identity / pose / clothing / brand / framing] unchanged."` Then state the change in one clean sentence. Models honor what's stated up front; tail-end preservations get ignored.

This rule serves simultaneously as a user manual and a model behavioral insight (tail-end preservations get ignored), converting empirical knowledge of how the model parses prompts into strict prompt engineering discipline. It shares the identical formulation with Route 1 of `image-edit`, demonstrating their shared lineage.

Localized edits are achieved via spatial language: Concrete spatial scopes like 'background only', 'the left object', or 'upper-right corner' are strictly respected by the model, whereas vague directives like 'make it more X' lead to drift.

nano-banana-edit/SKILL.md

**Localize with spatial language.** "background only", "the left object", "the upper-right corner", "above the headline" — concrete spatial scopes are honored. "make it more X" is vague and drifts.

Three invocation examples illustrate background swapping, batch watermark replacement, and left-object removal—all concluding with spatial bounding phrases as direct demonstrations of this discipline.

Batch consistency discipline: Series edits must lock `aspect_ratio` + `resolution` and share the same prompt syntax, ensuring batch outputs appear as 'siblings' rather than remixes; complex edits must be decomposed into small iterative passes rather than executed as a single compound prompt.

nano-banana-edit/SKILL.md

**Batch consistency** — when editing a series, lock `aspect_ratio` and `resolution`. Use the same prompt grammar across the batch so each output reads as a sibling, not a remix.

Combined with the 'Iterate small' section (`pass 1 changes background only, pass 2 swaps the subject's outfit`) and an anti-pattern list (compound instructions accumulating drift, passive voice, missing preservation targets, aspect ratio mismatches), this forms a comprehensive operational guideline.

Boundaries and routing: Multilingual in-image text edits are explicitly routed to GPT Image 2 edit; high-fidelity single-reference localized edits are routed to Flux Kontext; scratch generation is routed to the sibling nano-banana-2 t2i—the skill does not claim to be all-purpose.

nano-banana-edit/SKILL.md

| Edit multilingual text inside the image (signs, labels) | GPT Image 2 edit |

The Limitations section reinforces: `For multilingual in-image text edits, GPT Image 2 edit wins.`—establishing a coordinated division of labor with repository skills like `image-edit` (which includes the GPT Image 2 route) and the branded `gpt-image-edit`.

Execution delegated to runcomfy CLI + cloud inference: Executes `runcomfy run google/nano-banana-2/edit --input '{...}' --output-dir`, where the CLI sends a POST request to the full API URL, polls for completion, and downloads the results.

nano-banana-edit/SKILL.md

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

The input for editing tasks consists of external image URLs—the images are fetched by RunComfy's remote servers, while the local CLI transmits only JSON payloads.

2Core Capabilities

01Consistent batch editing for 1–20 input images (SKU galleries, A/B testing, character sheet variants)

02Identity-preserving background and outfit replacement (preservation-first syntax)

03Localized additions, deletions, and modifications guided by spatial language (left object / corner / above the headline)

04Batch framing lock + identical prompt syntax for producing sibling asset series (e-commerce brand asset recomposition)

05Ad creative A/B variants via seed locking + number_of_images

06Targeted in-image text replacement (altering text/color palettes while preserving composition)

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 图生图编辑端点（RunComfy Model API 托管）
network	生成结果下载白名单
network	用户提供的 1–20 张输入图 URL——由 RunComfy 模型服务器抓取
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: Declared external network access: model-api.runcomfy.net/v1/models/google/nano-banana-2/edit — Edit request submission (via runcomfy CLI)
本地工作目录: File I/O and artifact output: The CLI downloads generated results into the local directory specified by --output-dir (path determined by the caller)
SKILL.md: Declared invocation of host CLI interpreter or external runtime environment

Audit Notes: Evaluated under the unified tiering criteria: At runtime, it invokes the third-party npm package/CLI `@runcomfy/cli`. Each invocation reads user RunComfy credentials (`~/.config/runcomfy/token.json` or `RUNCOMFY_TOKEN` environment variable)—matching 'calls third-party npm/CLI and may read quotas/secrets -> orange'. The entire execution plane (CLI, Model API, token retrieval) resides outside this repository; documented claims (0600 permissions, 2 GiB limit, absence of shell expansion) cannot be verified in this pin, triggering the 'when uncertain, assign the higher tier' rule, classifying it as Orange. Furthermore, outbound network transfers exist: 1–20 input image URLs are fetched by RunComfy servers for inference (transmitting potentially unreleased SKU materials/faces to a third-party cloud in batch scenarios), and prompts are submitted via model-api.runcomfy.net; the documentation explicitly acknowledges image prompt injection risks. The skill itself consists purely of markdown with no scripts, obfuscation, or hidden endpoints, thus avoiding the Red tier.

5Second-Pass Verification

[ok] 端点 ID 与完整 URL 一致（google/nano-banana-2/edit） — description、schema 标题、How-it-works 完整 URL 三处一致，无拼写漂移。
[ok] 1–20 输入与 1–4 输出约束在 description/schema/Limitations 三处一致 — description（up to 20 inputs）、schema（1–20 / 1–4）、Limitations（1–20 input images / 1–4 outputs）数字自洽。
[ok] 与 image-edit Route 1 的同源关系 — preservation-first 话术、cyberpunk 背景与 AURA 水印示例在 nano-banana-edit 与 image-edit 中逐字一致——确认为同源内容，非本文件独有。
[ok] 多语种文字编辑让位 GPT Image 2 / 单参考让位 Flux Kontext 的导流 — When-to-pick 表与 Limitations 两处一致；对应品牌技能 gpt-image-edit/、flux-kontext/ 在仓库内真实存在。
[unlocatable] SKILL.md 对 CLI 行为的断言（token 0600、2 GiB、无 shell 展开） — CLI 源码不在本仓库，仅文档自述，无法独立证实。
[unlocatable] 『verified to produce strong results』sample prompts — 样例提示词自称 verified，仓内无评测证据可复核。
[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）。

6Verdict & Scenarios

Outstanding batch-consistent editing capability (1–20 images/call + primary first image + locked framing), making it the only i2i skill in the repository to treat batch processing as a first-class citizen

Preservation-first syntax backed by a mechanistic explanation (the model prioritizes front-loaded preservation targets), providing actionable guidance rather than vague advice

Precise spatial language syntax (background / left object / top corner) ensuring predictable localized edits

Honest capability boundaries: Explicitly yields multilingual text editing and single-reference high-fidelity modifications to more specialized skills

Transparent contract with complete API URLs and comprehensive security disclosures (including image injection risk warnings)

Best for: Agents and users with a RunComfy account performing e-commerce SKU batch background swaps, text replacements, watermark removals, ad creative A/B testing, and identity-preserved background replacements; batch consistency is its primary strength (20 images/call + sibling-grammar); also fits 'generation-to-editing' pipelines for fine-tuning assets initially generated by nano-banana-2.

Not for: Not suitable for multilingual (non-Latin script) in-image text modifications (should route to GPT Image 2 edit); not suitable for single-image, single-instruction high-fidelity localized editing (Flux Kontext is more specialized); not suitable for text-to-image generation from scratch (handled by nano-banana-2 t2i); not suitable for users without a RunComfy account, those who cannot upload assets to cloud infrastructure, or those needing to process untrusted external images.

Install Agent-Ready

① Verified Mirror 2026-09-09

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

sha256: ba0f1dd15b2c6baf…

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.