并发提交｜Vidu Q4 Preview 图生/参考视频（4 合 1）

Video generation is where a fan-out stops being a luxury. A Vidu Q4 clip takes minutes, not seconds, and if you're rendering six shots for a sequence you do not want to watch them happen one at a time. This node is Vidu_Q4_Preview_Video with a future output: same model dropdown, same seconds/resolution/ratio/generate_audio/is_rec/watermark controls, same fifteen image slots and three audio inputs - it just enqueues and returns instead of sitting there.

Why batch this specific node

Because the four Q4 variants are slow and expensive, and a batch lets you do the sensible thing: run the whole set cheap first. Six branches at 3 seconds and 540p tells you which prompts produce motion you like. Then re-run the two that worked at 1080p and a real duration. Sequential execution makes that loop painful enough that people skip it and pay for the expensive version of a prompt that was never going to work.

The receiver side is 并发接收视频（10 路）: ten future inputs, ten video outputs, plus status_json. The pack's video pool defaults to 10 workers, so slots and threads line up - dial it with SEEDANCE_VIDEO_CONCURRENCY (1 to 10) if you'd rather not have ten renders in flight.

export SEEDANCE_VIDEO_CONCURRENCY=4   # restart ComfyUI after changing it

One receiver for a mixed batch, incidentally. The receiver is generic over video futures, so you can wire a Vidu generation into future_1, a Topaz upscale of yesterday's clip into future_2, and get video_1 and video_2 out in order. Results are re-ordered by slot, not by completion time, so a fast clip can't shuffle your outputs.

Preflight is where you learn things

The wrapper runs the base node's validation synchronously before enqueuing, and this node's validation is unusually picky - which is a feature at batch scale. In particular:

• An R2V model with an empty prompt fails immediately. Good. That's a remote job you didn't pay for.
• I2V requires exactly one image. Two images in I2V mode is an error, not a coin flip.
• Same for R2V with zero images.
• An audio slot with both a local audio1 connection and an audio_url1 string fails, per slot.
• A malformed audio_url (anything that isn't HTTP or HTTPS) fails before anything is uploaded.

Ninety percent of your batch debugging happens in milliseconds, locally, with a message that names the input. That's genuinely better than the "one node in a chain of thirty went wrong somewhere" experience you get from most API packs.

Audio across a batch

generate_audio is per branch, and that's a nicer setup than it sounds. Branches 1–6 with audio on to see what the model invents, branches 7–10 with it off if you're scoring the sequence yourself with audio1/audio2/audio3 - the local AUDIO inputs get encoded to MP3 and uploaded automatically, so you can wire an audio file from the same graph without leaving ComfyUI.

Keep in mind that a ten-branch R2V sweep with reference audio is ten uploads per audio slot. It'll work; it just won't be instant, and rate limiting during the upload phase is the slowest part of a big batch.

failure_mode, and why it matters more for video

raise stops the run and names the slot. placeholder keeps the nine clips that worked, puts a placeholder video in the slot that didn't, and writes a redacted per-slot status into status_json. On a ten-wide video sweep, placeholder is usually the correct choice - a single upstream hiccup on slot 7 shouldn't cost you nine renders you've already paid for. Branch-level skip_error composes with both.

And note what "preview" means for a batch: if the provider's endpoint starts 502-ing at submit time, it does it for all ten. The pack's own changelog has a chapter on exactly that with Vidu Q3. Build the sweep, but test with one branch first.

Install

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

ComfyUI Manager: search ComfyUI Seedance, install, restart. One dependency, requests>=2.28, and the receiver node ships in the same pack. Load examples/并发视频2路最小验证.json first if you've never wired a submit/receive pair - it's two branches and it makes the pattern obvious. Each submit node still needs a key: its own api_config connection, or SEEDANCE_API_KEY in the environment, or config/.env.

Last thing, and it's the one that matters: ten slots means ten billed renders, several of them minutes long. The 2-way example exists for a reason. Draft the whole batch at 540p, then spend real money on the two branches that deserve it.