并发提交｜Zhenzhen Video V3.1

The Zhenzhen Video V3.1 concurrent submit node is the pack's fixed-duration video batcher: every model in it renders 8 seconds, and your only real choices are which of the three variants and which reference images. The model dropdown:

• zhenzhen-video-v31-fast (default) - the speed tier, up to 3 reference images.
• zhenzhen-video-v31-quality - the quality tier, up to 2 reference images.
• zhenzhen-video-v31-lite - the minimal tier: no images at all, text-only.

That per-model image cap is the thing to respect. The fast model takes three, quality takes two, lite takes zero, and the node will reject the wrong combination. This is the pack's clearest example of "the model you pick defines the input surface," which matters when you're filling ten slots at once and don't want to discover slot 8 is invalid.

Shared inputs: prompt (up to 20,480 chars), seconds (fixed at 8, still shown as a control), resolution (720p/1080p/4k per the README), ratio (16:9 default, forwarded as metadata.ratio), negative_prompt (optional, forwarded), seed (-1 = random, non-negative forwarded to metadata.seed), plus image1 … image3, api_config, and skip_error. Output is the single future (SEEDANCE_VIDEO_FUTURE) into 并发接收视频（10 路）.

How it works

Same concurrent machinery as the rest of the family - preflight against the selected model (this is where the image-count rule bites), video thread pool (10 workers; SEEDANCE_VIDEO_CONCURRENCY), future returned - over the /v1/videos endpoint. References upload and submit as images[]; the receiver gathers all slots in parallel, in order, with per-slot skip_error and its own failure_mode. Mixing a V3.1 future with any other video future is fine.

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

• The image caps are per-model and enforced. Wire three images into a lite slot and the preflight stops you cold. Pick the model first, then count your references.
• 8 seconds is the whole duration. There's no 4-or-15 knob here; every clip is 8s. If your storyboard beat needs 4s, this isn't the node.
• Batch discipline. Ten parallel 8-second clips is a real bill. The fast tier exists for iteration; use it, and fix your seed per slot if you want comparisons that mean anything.
• future is not a video. Receiver node, then SaveVideo.

Standing fine print: metered per clip, expiring links, data leaves the machine. V3.1 is the shop's "proper video" family - a fixed 8-second format is actually handy when you're batching, because every slot lands the same length and the receiver output lines up cleanly.