Zhenzhen_nano_banana2_edit

Comfly_nano_banana2_edit ("Zhenzhen_nano_banana2_edit") is the edit-focused node for Google's Nano Banana 2 - the speed-quality hybrid in the Banana family - and the headline numbers are the fourteen reference image inputs and the 1K/2K/4K size switch. If you're doing multi-image reference work - "keep this character, this outfit, this background" across several shots - this is the node the pack intends you to use.

It's the younger sibling of Comfly_nano_banana (the general-purpose one with seven models). Here the model list is narrowed to the 2-line: nano-banana-2, nano-banana-pro, plus the -2k and -4k pro variants, and there's an explicit mode switch between text2img and img2img. The aspect ratio list is wide - including portrait/landscape and the letterboxy 21:9 and 9:21 - and image_size gives you 1K, 2K, or 4K output.

It's a Google-class node in the T8mars/Comfyui-zhenzhen pack, the "Comfly" API-wrapper family that calls the author's paid reseller service. The key goes straight into an apikey field on the node. Submit → poll → download, same as the rest of the family, and the changelog shows the author iterating on this exact node hard (an "AUTO" aspect mode that follows your input image, plus the 2K/4K backup models for when Google's 4K output degrades to 1K).

What you set

• prompt - required, the instruction.
• mode - text2img or img2img.
• model - which Banana 2 flavor.
• aspect_ratio - 12 options including auto.
• image_size - 1K, 2K, or 4K.
• image1–image14 - the reference images. For character consistency, that's a lot of context to hand the model.
• response_format - url or b64_json for delivery.
• seed (cache control), skip_error, apikey.

Outputs

• image - the result as an IMAGE.
• image_url - the hosted result.
• response - the raw API payload.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen
# restart ComfyUI

Or ComfyUI Manager → "Comfyui-zhenzhen", then paste a key from the author's site into apikey.

Where people get burned

The 4K white-image problem. The README is explicit: at 4K on the default token group, Google's congestion can return blank/white output, and errors 429/500 are load limits that don't bill you. The fix is on the website - edit your token, drop the default group, add the premium group (优质分组) - and at 2K the failure rate drops a lot. The other gotcha is scope: fourteen reference inputs is a temptation, and every image you throw in is context the model has to reconcile. Two or three well-chosen references beat fourteen muddy ones every time. It's a reseller proxy, too, so those reference images all leave your machine - keep that in mind before you feed it anything sensitive.