Concurrent Submit | zhenzhen-kling-video-lowprice

Kling is Kuaishou's video model, and it's the kind of thing you can't run locally - no weights, no local option, API only (external-api-nodes.md has the whole "when local isn't the answer" argument). This node is how you call it cheaply, in bulk. It's the Concurrent Submit twin of Comfly_kling_video_lowprice: same model, same API, same settings - but instead of blocking your graph while one 5-second render crawls, it hands the job to a shared video pool and returns instantly.

What "Concurrent Submit" means here

Every node in this pack whose category starts with zhenzhen/ gets an auto-generated sibling with a ComflyConcurrent_..._Submit name. That's this node. The mechanism is simple: image jobs go into a pool of 30 worker threads, video jobs into a pool of 10 (COMFLY_VIDEO_CONCURRENCY env var if you want to change it), and each submit node just drops your request in and hands back a task handle. The actual Kling call still runs - it just runs off the graph's critical path.

The catch is graph shape. This node's only output is a task of type COMFLY_VIDEO_FUTURE. It does not return a video. To get your renders you wire that task into the pack's ComflyConcurrent_Video_Await collector, which waits for up to 10 video tasks, restores the order you plugged them in, and hands you the actual videos (plus a status string with a per-slot JSON report). Leave tasks unwired and you've paid for renders nobody collects. The pattern is: a row of these Submit nodes → one Await node.

The inputs that matter

The input list is identical to the original node, deep-copied, so nothing new to learn:

• model - 19 choices covering the Kling lineup: kling-v3.0-std-t2v, pro, v3-turbo, v3-4k, the kling-o3 family, and i2v versions of each. Text-to-video models default; switch to an i2v model and the reference images become relevant.
• prompt / negative_prompt - multiline strings. Kling actually respects a negative prompt, which is rarer in closed video APIs than you'd hope.
• seconds - 5 or 10.
• ratio - adaptive, 16:9, 4:3, 1:1, 3:4, 9:16, 21:9. The 21:9 option is nice if you're cutting cinematic wides.
• image1–image4 - up to four reference images for image-to-video models.
• api_config - this is the cheap domestic (国内版) route, so instead of pasting a bare key you wire in the pack's zhenzhen-seedance2-low-price-api-settings node, which bundles base URL + key.
• seed - the tooltip spells it out: it's an execution seed for ComfyUI cache control, not something the API consumes. Fixed reuses a cached result; randomize forces a new run.
• skip_error - leave off; when on, failures silently return an empty placeholder instead of raising.

How to install

It's part of the T8mars/Comfyui-zhenzhen pack (ComfyUI Manager → search "Comfyui-zhenzhen", or cd ComfyUI/custom_nodes && git clone https://github.com/T8mars/Comfyui-zhenzhen then restart). No model downloads - everything runs on the author's servers, and requirements.txt is standard stuff (requests, aiohttp, numpy, Pillow, torch). You do need a paid Zhenzhen API token from their storefront; the domestic store is api.seedance.nz.

Common issues

The servers are overseas, so from China you'll get 443/timeout errors without a proxy - the README is unsubtle about this and also blames aggressive firewalls. Model availability and price also depend on which token group your key is in. And remember: a Submit node that errors will fail the whole batch unless you tick skip_error - for a big batch, that's often the more forgiving choice.