RH Nano Banana Pro

Nano Banana Pro is Google's closed image model - the one that nails photorealistic edits, renders 4K natively, and that you will never be able to download. There are no open weights, so the only way to get it inside a ComfyUI workflow is through an API node. That's exactly what RH Nano Banana Pro is: a single node that sends your prompt and image to RunningHub's hosted Nano Banana Pro endpoints, waits for the result, and drops the finished image back into your graph like a local sampler made it.

The name "RunningHub" is the tell. RunningHub is a cloud platform - a cheaper, China-based cousin of RunPod or Comfy Cloud - that hosts full ComfyUI workflows and also sells model APIs. This node talks to those APIs, not to Google directly: you pay RunningHub per call, no Google Cloud project required. It's one of the handful of ways people run the "generate the frame on Nano Banana, animate it with Wan" hybrid that's all over r/comfyui.

How it works

Under the hood it's a task-queue client, not a model. It:

1. Uploads your input image to RunningHub's media endpoint and gets back a URL - this is what flips it into image-edit (I2I) mode.
2. POSTs a job to either the third_party or official Nano Banana Pro endpoint (more below).
3. Polls the task status every 5 seconds, up to 60 attempts (five minutes of patience).
4. Downloads the result and converts it to a standard IMAGE tensor.

No image wired in? Then it's plain text-to-image, and it adds an aspectRatio to the request. The official edit endpoint rejects that field, which is why the node only sends it when you're in T2I mode.

The inputs that matter

• api_key - your RunningHub API key. Not a Google key, not an OpenAI key. Get it from RunningHub, paste it here. It's stored in plain text in the workflow JSON, so don't share workflows containing it.
• prompt - your instructions, plain English works well here (Nano Banana understands natural language far better than a checkpoint's prompt grammar).
• resolution - 1k, 2k, or 4k. The Pro model's whole selling point is native 4K, but 4K jobs cost more credits and take longer. Start at 1k to iterate.
• endpoint_mode - auto_fallback (default) tries the third-party endpoint first and falls back to the official one if it fails; third_party_only and official_only pin one. If one endpoint keeps erroring for you, try pinning the other.
• image (optional) - wire any image in and it becomes an edit/reference task.
• aspect_ratio (optional) - auto, 1:1, 3:4, 4:3, 16:9, 9:16. T2I only.

The outputs are image (a normal IMAGE tensor - straight into a preview, upscaler, or video model) and info, a JSON string reporting which endpoint served you, which mode ran, elapsed seconds, and usage/credit details when the API returns them.

Installing it

Easiest via ComfyUI Manager - search "comfyui_runninghub_nanobanana" (or "RunningHub NanoBanana") and hit install, then restart ComfyUI. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ThanaritKanjanametawatAU/comfyui_runninghub_nanobanana

Then restart. Good news on the heavy-dependency front: there are none. The actual requirements are just requests and Pillow. No model downloads, no VRAM, nothing to babysit.

Where people get burned

• It costs money, per call. Each generation is metered against your RunningHub balance. A draft session of fiddly edits can surprise you, so watch the info output for usage numbers.
• Your data leaves the machine. The whole point is that your prompts and reference images go to RunningHub, which relays them to Google's servers. If your workflow touches sensitive images, this is not the node for it.
• Content policy is not negotiable. RunningHub is a strictly no-NSFW platform, and Google's model is aggressively filtered anyway. You cannot prompt around it - closed weights mean no local patches.
• There's no seed and no caching. The node always re-runs on queue, so you can't reproduce a result exactly, and you'll pay again every time you re-queue.
• The README is the cookiecutter template. Seriously - the repo's docs are the boilerplate every new ComfyUI extension starts from; the source code is the real documentation. The pack is fresh (v0.0.2, single node). With any new API-key node, read it before you trust it - this one's behavior matches its description, but it's the category to be careful in.

If you just want Nano Banana Pro without a third-party key, ComfyUI's own Partner Nodes offer it on prepaid credits with Comfy's vetting behind it. This node's reason to exist is RunningHub's pricing, billing, and region access instead. For the classic workflow - clean start frame here, then Wan or FFLF for the motion - it slots in as your first node and does exactly one job.