Concurrent Submit | Zhenzhen_nano_banana_fal

The first-gen Nano Banana model, reached through fal's queue instead of Google's direct route, and wrapped in Zhenzhen's concurrent layer so you can run a pile of them at once. It's the modest cousin of the Banana 2 and Pro fal nodes - fewer features, cheaper vibes - but if you're batch-generating a lot of images where "good enough" is the bar, this is a workhorse.

The workflow shape is the same as every Submit node in this pack: you feed it a prompt and get back a COMFLY_IMAGE_FUTURE task, not an image. That task goes into Concurrent Collect Images (ComflyConcurrent_Image_Await), which pools up to 30 tasks in the shared image pool, waits for all of them, and emits image_1…image_30 plus a status JSON. Wire those into your save/preview nodes. That's the entire trick: submit everything first, collect once everything's done, never block the graph on a single generation.

The inputs

• prompt - the text prompt, multiline.
• model - nano-banana for text-to-image, or nano-banana/edit for image editing. The edit model is what charges you extra: Zhenzhen bills image uploads separately, and the README explicitly calls out that the /edit model "会产生额外的图片上传费用" (extra upload cost). Check your task log if the number looks higher than expected.
• num_images - 1 to 4 images per call.
• image1 … image4 - reference images (used when model is the edit variant).
• image_way - image vs image_url, an older enum than the newer nodes' image_url/base64. image uploads your tensor to their server; image_url lets you pass a URL directly.
• seed - fal caps it at 65535; 0 = random.
• apikey and skip_error - the usual.

Install

Same pack, same drill:

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

Or ComfyUI Manager → search Comfyui-zhenzhen → restart. Get a token from ai.t8star.org (overseas) or api.seedance.nz (domestic, ~7.5–8折), paste it into apikey, and remember this pack calls out to a reseller's server - your prompts and images leave the machine.

Gotchas

• Pre-deduction billing. fal lanes deduct an estimate up front and refund the difference after the job settles. Don't panic at the initial charge; the README walks through exactly this.
• Polling timeouts. The defaults give you an hour (max_poll_attempts × poll_interval). If it times out, the job may have completed server-side - pull it from the async-task page before resubmitting.
• Upload costs add up in edit mode. If you're doing thousands of edits, the upload fees can rival the generation fees. That's the price of convenience, not a bug.