JM Gemini Image Generator

If you've been jealous of everyone getting clean, coherent Nano Banana images while you're stuck fighting your SDXL checkpoint's hands, this is the node for you. JMGeminiImageGenerator puts Google's Gemini image models - gemini-3-pro-image-preview (that's Nano Banana Pro) and the faster gemini-2.5-flash-image (Nano Banana) - right into a ComfyUI graph as a single node. You prompt it, it talks to Google, and an IMAGE tensor comes out the other side ready to plug into your normal preview/save/upscale chain.

The whole point is keeping your ComfyUI workflow while offloading generation to the cloud. Zero VRAM for the actual image gen - your GPU just renders whatever post-processing you chain after it. That's a big deal if you're on a 6GB card, and it's the reason this category of "cloud model as a node" pack keeps growing.

How it works

Under the hood it's a thin wrapper around the official google-genai SDK (client.models.generate_content with response_modalities=['TEXT', 'IMAGE']). You give it an API key, it calls Google, and it hands back the first image in the response as a ComfyUI tensor - plus an auto-saved PNG in ComfyUI/output/ named something like gemini3pro_text2img_1234567890.png.

Three modes, decided by which image inputs you fill:

• Text-to-image - leave all image slots empty.
• Image editing - connect exactly one image; your prompt becomes an edit instruction ("add a sunset background").
• Image-to-image - connect 2–10 images; the model combines or transforms them per your prompt.

The inputs that actually matter

• gemini_api_key - from Google AI Studio. Without it nothing happens. This is pay-per-use, not free: KB pricing puts Nano Banana Pro around $0.04–$0.24 per image depending on resolution, and it adds up fast if you iterate.
• prompt - the prompt. In edit mode, the instruction.
• model - gemini-3-pro-image-preview (default, higher quality, supports 1K/2K/4K) vs gemini-2.5-flash-image (faster and cheaper, ignores the resolution dropdown entirely and maps aspect ratio to fixed pixel sizes).
• aspect_ratio - one of ten (1:1 through 21:9).
• resolution - 1K/2K/4K, only honored by the Pro model. Set it on Flash and wonder why nothing changed - that's expected.
• image1..image10 - optional IMAGE inputs for edit/i2i modes.

One trap worth knowing: seed does nothing. The source literally comments that the seed is only used for ComfyUI re-execution and never passed to the API. Don't expect seed-locked reproducibility - if you want a different take, change the prompt or re-run.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/synthetai/ComfyUI-JM-Gemini-API
cd ComfyUI-JM-Gemini-API
pip install -r requirements.txt

Then restart ComfyUI. Easier: search "ComfyUI-JM-Gemini-API" in ComfyUI Manager and let it handle it. The one dependency that actually matters is google-genai==1.52.0 - it's pinned, and httpx[socks] is there for proxy users (set HTTP_PROXY/HTTPS_PROXY env vars if you need them, which most users in China do). No model files to download.

Common issues

• "Gemini API key is required" - the key field is empty or pasted with whitespace.
• "No images were generated" - usually a prompt that tripped Google's safety filters, or a key with no quota. This model is heavily censored (the KB quotes the community's "powerful, unimaginative, massive hidden biasing"); if your prompt is even close to a blocked subject, you'll get text back instead of an image.
• Resolution changes do nothing - you're on Flash. Switch to Pro.

Pair its output with the pack's JM Gemini Watermark Remover and your pipeline is complete: generate, strip the corner logo, save. If you want Gemini quality without paying per image, the pack's other node - the Reverse Engineering one - is the "free" cookie-based path, with caveats of its own.