图片｜Gemini 14图极限生图

"Gemini 14图极限生图" - 14-image extreme generation. That's the whole pitch, and it's a real one: TikpanGeminiImageMaxNode is the Gemini image node that lets you wire in fourteen reference images and have the model fuse them into one result. For multi-reference work - a brand sheet, several product angles, a character across outfits - that's the top of the range this pack offers, and it's why the node exists next to the simpler GPT-Image nodes. It runs on gemini-3-pro-image-preview (highest quality, default), gemini-3.1-flash-image-preview, or the nano-banana family.

How it works

It's a sync image call: your 修改指令 (which doubles as the generation prompt), model pick, resolution tier, aspect ratio, and up to 14 reference images get sent to Tikpan's Gemini endpoint, and the result comes back as 🖼️_生成结果图 (IMAGE) plus 📄_渲染日志 (STRING). The 调用方式 field switches between three relay endpoints - gemini 原生 (default, recommended), images_generations, or chat_completions - for compatibility with older workflows. The node hides a seed parameter that the author hadn't confirmed was honored upstream, so don't expect reproducible outputs across runs.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 修改指令 - what to generate or edit. The default asks for a vertical poster "referencing the provided images, subject consistent, cinematic lighting" - a good template for how the multi-ref mode works: describe the fusion, let the images carry the identity.
• 模型 - pro (best quality), flash (faster), or nano-banana variants. The tooltip's guidance: pro is highest quality, flash is faster, nano-banana is the most stable for multi-image fusion - which is a hint about where multi-ref runs can get flaky.
• 分辨率 - none (model decides) / 1K / 2K / 4K. Higher = slower and pricier.
• 画面比例 - six options, default 9:16 (short-video friendly).
• 参考图_1 … 参考图_14 - the feature. Wire images in and the model fuses them.

Installing it

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

restart (or ComfyUI Manager → "Tikpan"). No models, no heavy deps - relay + image decode/save.

Where people get burned

• 14 references is the ceiling, not the sweet spot. More references means more tokens and more ways for the model to drift. Start with 2–4 for identity work; reserve the full 14 for deliberate mood-board fusion. The tooltip's "nano-banana 系列多图融合更稳" is the author's own hint that the pro model isn't always the right choice for heavy multi-ref.
• Seed isn't guaranteed. The schema notes the seed parameter is hidden because it wasn't confirmed to work upstream. Don't build reproducibility expectations on it.
• Cost scales with references and resolution. 4K + 14 refs is the most expensive combination in the image section. 2K is the sane default.
• 修改指令 vs prompt confusion. There's no separate "prompt" field - this one field does generation and editing. Give it the full brief.

If your image work is single-shot generation, the simpler GPT-Image-2 nodes are cheaper and less fiddly. If you're fusing a dozen references into one canvas, this is the node. It's a specialist tool, and it knows it.