Gemini3-image-Nano

Nano Banana is Google's closed image model and the strongest of its class as of 2026 - and you cannot run it locally, period. This node is the "full-blooded" (满血) Gemini image wrapper in the pack: instead of the trimmed-down single-purpose editor, it exposes nearly the whole parameter surface - which model, how hard it thinks, how safe the output is, what resolution - plus optional multi-turn context so an image can be edited across several turns of conversation. If your workflow wants Gemini image generation with the knobs actually exposed, this is the node.

How it works

The node POSTs your prompt (plus optional input_images and a context) to the pack gateway, selecting the backend model from its dropdown: Gemini 2.5 Flash Image (the original Nano Banana), Gemini-3-pro-image-preview, or Gemini-3.1-flash-image-preview. The interesting mechanics are the two outputs besides the image: a text output and a context output. The model can return text alongside the image (Gemini 3 explains or captions what it made), and context carries the conversation state onward - wire it into the pack's ContextNode or into another Gemini node to keep the thread alive. That's the multi-turn edit loop: image in, edit instruction in, context in, next image out.

The inputs that matter

• model (enum) - which Gemini image model. Note the pack's own note: Gemini 2.5 Flash Image doesn't support 2K/4K output; the Gemini 3 models do.
• thinking_level (minimal/low/medium/high) - how much reasoning before generating. High is better text rendering and composition at the cost of time and money.
• safe_level (high/medium/low) - output safety filtering. Low is less filtered (still Google-filtered - nothing here bypasses the model's own moderation).
• media_resolution (Default/Low/Medium/High) - how hard the model looks at your reference images.
• resolution (1K/2K/4K) and aspect_ratio (ten presets) - output framing.
• System_prompt - a system instruction, if you know how to use one.
• Web_search (BOOLEAN) - lets the model search the web as part of generation (a genuinely unusual thing to see on an image node).
• seed (INT) - reproducibility.

Optional: prompt, input_images (batch), context. Outputs: image (IMAGE), text (STRING), context (ANY).

Installing it

Part of comfyui-MJAPI-party. ComfyUI Manager → "mojieapi_party" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

Register at mojieaigc.com, put your key in config.ini (leave BASE_URL = https://www.mojieaigc.com/v1/completions as shipped), restart:

[API]
KEY = your_api_key
BASE_URL = https://www.mojieaigc.com/v1/completions

No models to download.

Common issues

• "1K" output when you asked for 4K. Check the model - the 2.5 Flash Image backend caps at 1K in this wrapper. Switch to a Gemini 3 model for 2K/4K.
• Multi-turn edits aren't "sticking." The conversation lives in context. If you aren't feeding the previous turn's context back in (or reading it from ContextNode), each run is a fresh, contextless generation.
• Cost creeps with the knobs. High thinking + 4K + web search is the premium combination; on a metered key it's the difference between a handful and a full session. Iterate on low thinking at 1K, then spend on the final pass.
• Error image with text - the pack's standard failure display (key/balance/timeout/refused).