图片｜Nano Banana Pro

Nano Banana Pro is Google's flagship closed image model (Gemini 3 Pro Image) - 4K-native output, up to 14 reference images, the strongest text-in-image rendering in the closed world, and absolutely no open weights. The official route to it is a Google Cloud project, Vertex AI, and a billing account. This node is the reseller route: same model, Tikpan's sk- key, one field, done. If the friction of Google Cloud is what's kept you from Nano Banana Pro, this is the pack's headline argument.

It's worth being clear about what you're buying, because the reseller layer is the entire story. Tikpan relays your requests to Google's Gemini image API (this node supports both gemini原生 and openai兼容 protocols). You get the model's quality - the 14-reference fusion, the 4K output, the excellent text rendering - plus Google's moderation (Nano Banana is among the most aggressively filtered closed models, per the KB; the January 2026 policy tightening didn't loosen it). Cost is per image and resolution-gated. The reseller's advantage is cost, region access, and billing simplicity; its disadvantage is that you're trusting a third party with your prompts and a key. Same tradeoff as every node in this pack, just with the most famous model in it.

How it works

修改指令 is your prompt (the field name betrays the author's default use-case - modifying/generating product imagery). The 模型 dropdown offers gemini-3-pro-image-preview (quality, slower) and gemini-3.1-flash-image-preview (fast, cheap). 分辨率 runs 1K/2K/4K/none, and the node does the pixel math itself - picking 4K plus a 16:9 ratio produces a properly sized request instead of leaving it to the model. 画面比例 covers the usual ratios up to 21:9 and 2.35:1.

The 14 参考图_N ports are the party piece: each is a visual reference the model fuses into the generation, which is how you get "keep this product's identity across a scene" without LoRAs. Outputs are 🖼️_生成结果图 (IMAGE) and 📄_渲染日志 (STRING).

The inputs that matter

• 修改指令 - the prompt; specific English wins, and it's fine to be long.
• 模型 - pro for deliverables, flash for drafts.
• 分辨率 - 2K is a good default; 4K when the file actually matters.
• 参考图_1–参考图_14 - references. Start small; every extra image costs upload time and can dilute the output.

Install and first run

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

Restart (or Manager → "Tikpan"), paste your sk- key, write a prompt, run. First result is synchronous and returns in seconds. No downloads, no Google Cloud setup - that's the whole value proposition.

Common issues

The pack-wide error family (401/402/429) applies, plus the ones specific to image calls: big reference uploads stalling (compress or host them), and the occasional request that Google's own safety filter just refuses - which no reseller can route around, since the filter lives in the model, not the node. On the openai-compat protocol, keep 返回方式/payload defaults aligned with the endpoint; if you see payload-format errors, the 调用方式 dropdown is the first thing to flip.