Superside Nano Banana Pro Edit Node

Nano Banana Pro is Google's context-aware image editing model, and "context-aware" is the whole reason people hunt for it in ComfyUI: you hand it your image plus a plain-English instruction - "make it golden hour, keep the product unchanged" - and it edits the whole frame with a real understanding of what's in the picture, instead of the blunt img2img blend you'd get from a local sampler. It's a closed model: there are no weights to download, which means an API wrapper like this one is the only way it shows up in a ComfyUI graph.

This node calls fal-ai/nano-banana-pro/edit with up to six reference images. The "up to six" part is the pro-level feature - you can feed it the product, the model, the prop, and the lighting reference all at once and edit them as a scene, not as a single image. That's a genuinely different thing from the single-image edit nodes.

The inputs that matter

• prompt, image_1, api_key - the essentials. The prompt is the edit instruction; the model's whole pitch is that it follows intent, not just keywords.
• image_2 through image_6 (optional) - extra references. More context, more scene-level editing. They're only included in the call if connected.
• resolution - 1K (default), 2K, or 4K. This is the headline dial: 4K edits are the reason you came here, but they're the expensive end. Build at 1K, upscale at the end for keepers.
• aspect_ratio (default auto) - auto inherits your input's ratio; the long list (21:9 down to 9:16) is there if you want a specific crop instead.
• num_images (1–4) - variations per prompt. Each one is a billable call.
• output_format - jpeg, png, or webp. PNG for anything you'll touch again.
• sync_mode (default off) - leave it.

Two outputs: images (the edited IMAGE) and description (STRING) - a text summary of what the model says it did, which is genuinely useful for logging or deciding which variation kept your intent.

How it works

Same Superside/fal pattern as the rest of this pack: upload your images, call the endpoint, download the result. Google's model does the thinking on their servers. The honest trade is the API-node trinity: per-call cost (video-adjacent, but 4K edits aren't free), data leaving the machine, and a moderation filter you can't bypass.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes
pip install -r requirements.txt

Restart ComfyUI, find it under Superside, and paste a fal key into api_key (blank falls back to a FAL_KEY env var).

The trap nobody mentions: color drift

Edit a clean original, and the model's re-render usually sneaks a warm or red cast in - the Superside README calls this "generative color drift," and it accumulates if you run an image through a Nano Banana pass repeatedly. The pack ships a White Balance node (calibrate from a neutral patch) and a Color Match node (pull the drifted edit back toward the original) precisely because of it. If your Nano Banana results keep coming back subtly warmer than the source, that's not you - that's the model - and the fix is a local color-correction node downstream, not more prompting.