并发提交｜Zhenzhen Video GK v1.5

Zhenzhen Video GK v1.5 is the video half of the shop's GK family, and this node is its concurrent submit wrapper. One model (zhenzhen-video-gk-v15), prompt-driven, with optional reference images. If the GK image nodes are the "make a still, cheaply" tier, this is "make a short clip, moderately cheaply," and the concurrent version exists so a batch of prompts can all render at once instead of serially waiting on the gateway.

The input set is small and worth memorizing because it recurs across the Zhenzhen video nodes:

• prompt - up to 20,480 characters.
• seconds - duration, submitted as a string; 6 is the default.
• resolution - 720p default.
• ratio - aspect ratio, forwarded as metadata.ratio; 16:9 default.
• negative_prompt - optional, forwarded to metadata.
• seed - -1 for random; non-negative values are forwarded to metadata.seed, so it's a real seed, not just a cache key.
• image1 / image2 - optional reference images, submitted as images[].

Output is the single future (SEEDANCE_VIDEO_FUTURE) into 并发接收视频（10 路）.

How it works

The standard concurrent pattern: inputs validated up front, job handed to the video thread pool (10 workers by default; SEEDANCE_VIDEO_CONCURRENCY to dial down), future returned immediately. References upload, the request is built on the /v1/videos endpoint with the metadata structure above, and the receiver collects all ten slots in parallel, in order, with per-slot skip_error and its own failure_mode. Because the receiver is generic, you can mix a GK v1.5 future with, say, an upscaler future on the same run.

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

• Video is the expensive medium. Ten parallel clips is ten paid renders, and video metering dwarfs image metering. Iterate at 720p/6s; this is a batching node, which makes the temptation to fill all ten slots strong and the bill equally strong.
• Seed discipline. Since seed is forwarded, -1 on every slot gives you ten random takes you can't meaningfully compare. Set fixed seeds per slot when you're testing prompt wording.
• Reference behavior. Connect image1/image2 and the model animates from them; leave them empty and it's pure text-to-video. Neither is wrong, just know which you asked for.
• future → receiver, not SaveVideo. The family rule.

Standing fine print applies - metered per render, expiring result links, prompts and reference images leaving the machine. For "spin up a handful of mid-tier clips in parallel," this is a perfectly boring, dependable node, and in the API-video world boring is a feature.