并发提交｜Zhenzhen Image G 图像生成/编辑

"Zhenzhen" is the brand of the shop that runs this whole API (贞贞的平价AI小铺 - roughly "Zhenzhen's budget AI shop"), and the Zhenzhen Image G family is its own line of image models. This concurrent submit node wraps the three-model Image G node: the default zhenzhen-image-g-v2-lowprice (the budget tier, and the one the node defaults to), plus the two G-2 models, zhenzhen-image-g2-t2i and zhenzhen-image-g2-i2i. Its job is to let you fire off a batch of these cheap images in parallel instead of one at a time.

The model choice splits the inputs in two, and knowing which fields apply where is half the battle:

• zhenzhen-image-g-v2-lowprice (default) - the budget workhorse. Resolution 1k/2k/4k, its own size dropdown (common ratios plus a custom option where you type 5:4 or 2048x1024 into custom_size), and n for 1 to 10 images per request. Up to 16 reference images, or none at all.
• zhenzhen-image-g2-t2i - the plain text-to-image G-2. Resolution is fixed at 1k, uses ratio instead of size, no n.
• zhenzhen-image-g2-i2i - the G-2 edit model. Same fixed 1k + ratio, but it requires reference images (1 to 10 of them; only the first 10 of the 16 slots are used).

Prompt limits differ too: Lowprice is 5–5,000 characters, G-2 up to 20,000. The node preflights all this before the job goes anywhere, so a bad combo fails fast at the submit node rather than halfway through a batch.

Output is the single future (SEEDANCE_IMAGE_FUTURE), wired into 并发接收图片（30 路）.

How it works

Standard concurrent architecture: validate up front, drop into the image thread pool (up to 30 workers, tunable via SEEDANCE_IMAGE_CONCURRENCY), return the handle immediately. Reference images upload in slot order, the request is assembled per-model, and the receiver collects the finished images in slot order. skip_error per slot; failure_mode on the receiver decides raise-vs-placeholder. You can mix this node's futures with any other image submit node's - the receiver doesn't care what generated each slot.

Installing it

ComfyUI Manager → search ComfyUI Seedance, or:

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

Restart, look under Seedance/并发提交. requests and an api.seedance.nz key; nothing to download.

Where people get burned

• Assuming every model takes every field. ratio silently does nothing on Lowprice (it uses size), size does nothing on G-2, n doesn't exist on G-2 at all. Pick the model first, then read which knobs are live.
• Lowprice is the budget tier for a reason. It's the default, and it's cheap - great for drafts and batch experiments. When you're happy with a composition, switch to G-2 or a higher-res model rather than cranking Lowprice to 4k and wondering why it's still budget-flavored.
• The future isn't an image. It goes to the receiver, never straight into Save Image.

Standing fine print: per-image metering (all the more reason the Lowprice default exists), expiring result links, and prompts/refs leaving the machine to the api.seedance.nz endpoint.