Modelverse Gemini Flash Image

"Nano Banana" sounds like a joke node name, but it's Google's actual branding - it started as an anonymous codename on a leaderboard, Google execs started posting banana emojis, and the name stuck. Under it sit Google's Gemini-native image models: Nano Banana is Gemini 2.5 Flash Image, and Nano Banana 2 is the newer Gemini 3.1 Flash. The Modelverse NanoBanana node gives you both through UCloud's cloud API, so you get Google's image generation in your ComfyUI graph without a Google Cloud account, a GPU, or a single local model file.

This is the node to reach for when prompt-following and weird, creative compositions matter more than local control. The KB's closed-source notes are worth keeping in mind: Gemini image models are technically strong but heavily filtered, and Google tightened the safety dials further in 2026 - famous IP and celebrity stuff will get refused, and you should expect the censorship you'd get from Google, because it's Google.

The inputs that matter

• client - the pack's standard wiring from Modelverse Client or Secret Client.
• model - gemini-3.1-flash-image (default, the newer Nano Banana 2) or gemini-2.5-flash-image (original Nano Banana). Start with the 3.1 - it's the speed/quality hybrid.
• prompt - descriptive text. The default is a demo about a "nano banana dish," which tells you everything about where the name came from.
• mime_type - image/png or image/jpeg for the output. PNG for quality, JPEG if you're batch-generating and want smaller payloads.
• num_requests (1–10) - how many separate generations to fire. Unlike the Qwen node there's no num_images; each request is one image, and requests run sequentially here rather than as one concurrent batch.
• image (optional) - plug in an IMAGE and the node switches from text-to-image to edit mode: it sends your image as part of the prompt context. That's the "add this to the picture" trick Gemini is famous for.

The single image output is a batched tensor - one per request - so num_requests of 3 gives you a batch of 3.

How it works

The node calls Modelverse's Gemini-style endpoint (/v1beta/models/{model}:generateContent) and parses the response for inline image data, which Gemini returns base64-encoded rather than as a URL. If you ever see a blank result with a warning in the console, that's the model refusing to return an image (usually a safety refusal) - the node logs it and moves on.

Install and gotchas

One install covers the pack: ComfyUI Manager → "ComfyUl-UCloud" → Install → restart, or git clone https://github.com/ucloud/comfyui-plugin into custom_nodes. Nothing heavy to download.

Two things to budget for: it's a metered API, so num_requests is a bill multiplier, and Google's image models tend to be slower per image than the diffusion-style endpoints. And the SynthID watermark thing is real - Google embeds invisible watermarks in these images, so if that matters for your use case, this isn't the node for you.