Zhenzhen Image NB Pro / 2.1 Lowprice 生成/编辑

Two model names sit inside this one node - zhenzhen-image-nb-pro-lowprice and zhenzhen-image-nb-2.1-lowprice - and the "lowprice" in both is doing a lot of work. What you're actually renting is Google's Nano Banana Pro / Nano Banana 2.1 image model, routed through the Seedance reseller gateway instead of Google's own metered API. Same family that gives you 4K output and up to 14 reference images (closed-source-models.md); different till, looser billing, and quality that can wobble compared to the first-party route. That trade is the whole reason this node exists, and it's worth knowing what you're getting before you build a workflow on it.

The 14 references are the headline feature and not a gimmick. Fourteen ordered images is enough to hold a character's face, a product, a set of text labels and a style board at once - precisely the job local pipelines solve with LoRA training, IP-Adapter stacks or a pile of masking. Here you just connect them. No training run, no VRAM.

What the node sends

The request body is short: model, prompt, n fixed at 1, a top-level size (the aspect ratio), and metadata.resolution. If any image1…image14 slot is connected, those images get uploaded one at a time as PNGs and appended as an ordered images array - ordered, meaning slot number is the order the model sees them, so don't reshuffle.

No reference images connected means text-to-image. Any reference connected means edit. There's no mode switch to get wrong, which is a small mercy in a pack this big.

n being pinned at 1 is the thing to plan around: one image per run. There's no "give me four variations" here. If you want variations, either run it four times or - better - use the concurrent submit version of this node and fan out properly.

The inputs that matter

model picks the lane. prompt takes 1–5000 characters. resolution is 1k, 2k or 4k. size is where things get slightly interesting: the dropdown holds fifteen values, but the front-end swaps the list per model. Nano Banana Pro gets the standard set (auto, 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9); switch to 2.1 and you additionally get the extreme ones - 1:4, 4:1, 1:8, 8:1. Those four are genuinely useful for banners and phone-wallpaper crops and almost nothing else in ComfyUI's image-node ecosystem offers them.

The reference slots behave like the rest of the pack: only the next empty slot is visible, so the node doesn't start life as a fourteen-port monster. Connect image1 and image2 appears, and so on. Each slot takes exactly one image - a batch in a single slot is rejected with "each NB Lowprice slot requires one image."

Outputs are the standard quartet: image (wire to preview/save/upscale), image_url, task_id, and response for when you need to see what the API said. Like every generation node in this pack there's also a seed widget, and like the rest it's a cache key only - the value is stripped before the request and never reaches the model. Fixed seed means "don't re-bill me for an identical re-run," not "reproduce that image."

Install

Same pack, same two steps. Manager: search ComfyUI Seedance, or comfy node install seedance. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
cd .. && python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

Dependencies are literally requests>=2.28, so install is a few seconds and there are no weights to download. Restart, then give it a key via a Seedance API Config node, SEEDANCE_API_KEY, or config/.env.

Where people get burned

Model/ratio mismatch after loading a saved workflow. The ratio list is swapped client-side. Save a graph with 8:1 selected and a different model chosen, reopen it, and the node throws "unsupported NB Lowprice aspect ratio for this model" - clear enough as errors go, but it reads like a bug when it's just a stale dropdown value. Set the model first, then the ratio.

Fourteen uploads are slow. Each reference is uploaded individually before the job is even submitted, and the gateway rate-limits uploads per token. A full 14-image edit can take noticeably longer to start than it does to generate. The node retries and waits rather than failing, so it just looks stuck - watch the console, not the progress bar.

Empty prompts fail fast. Prompt validation runs before submission, so you get an error instead of a billed no-op. Good news.

And the honest caveat that applies to every node in this pack: you're on a reseller. If the image quality or the moderation behaviour doesn't match what you've seen from Nano Banana elsewhere, that's the routing, not you. Test a few prompts at 1k before committing a production workflow to 4k runs.