并发提交|Zhenzhen Motion Video 动效视频 (Standard / Pro)
The concurrent submit variant of Zhenzhen Motion Video
- api_config
- future
Why this exists
Zhenzhen_Motion_Video renders Canvas 2D motion graphics - kinetic type, intros, data stories - on a remote service. Your GPU does nothing; your wall clock does all the waiting. One 30-second render blocks the whole graph for minutes, and batching five variants of the same intro means five sequential waits for zero local compute.
The 并发提交| ("concurrent submit") family fixes exactly that. Same node, same parameters, same remote render - it hands the job to a background thread pool and returns immediately. This one is SeedanceConcurrent_Zhenzhen_Motion_Video_Submit; its partner on the other end is the 并发接收视频(10 路) receiver.
Internalise this first: every submission is a metered call to api.seedance.nz, the author's paid gateway. Concurrency isn't free throughput - it's ten bills in flight instead of one.
Same inputs, different control flow
The input panel is a copy of the original's: model, operation, prompt, duration 15/30/45/60, ratio, resolution, the ten template_id looks from hype-kinetic to cinematic-trailer, plus title, subtitle, keywords, signature, language, music, assets, and the revision pair revise_from / revision_instruction. api_config, skip_error and seed are there too.
The output is a single future of type SEEDANCE_VIDEO_FUTURE. One wire, and it doesn't carry a video - it carries a promise of one.
The pack builds this class programmatically from the original node, and one detail matters: the wrapper deliberately doesn't inherit the original's validation, because older ComfyUI builds would take one bad model choice and paint "invalid input" across every widget on the node. But it still runs the original's strict check synchronously before the pool, so a missing prompt or four assets that aren't plain public HTTP(S) URLs error instantly instead of eight minutes later at the receiver. With skip_error on, that failure becomes a pre-failed future instead, so one bad slot doesn't take the batch down.
Collecting results
Wire future into future_1 (up to future_10) on 并发接收视频(10 路), and put the receiver's video_1 … video_10 where the clips should go. It waits for all of them, then emits in slot order - slot 3's job always lands on video_3. Empty slots get a generated placeholder video rather than breaking the graph, so numbering stays stable.
failure_mode decides what a failure costs you. raise (default) stops at the first bad slot and names it: Concurrent video slot 5 failed: …. placeholder keeps going, drops an error video in that slot and leaves the rest intact. Either way status_json is a per-slot summary - state, source node, a redacted error string (keys, signed URLs and task IDs scrubbed), and which optional outputs the job actually populated. Running unattended, use placeholder and read status_json.
How many run at once is capped by the pack's video pool, 10 workers by default:
export SEEDANCE_VIDEO_CONCURRENCY=4
Images have a separate pool (SEEDANCE_IMAGE_CONCURRENCY), so a video batch never starves your upscales. Futures of the same type are interchangeable too - the README explicitly allows mixing this with other video submit nodes in one receiver: a Motion Video render in future_1, a Seedance clip in future_2.
The catch nobody mentions up front
The receiver only carries each job's primary output. For this node that's video. Nothing else survives the trip - no task_id, cover_url, preview_url, storyboard or fallback_scenes.
That bites twice. You can't chain a revise run off a concurrent render, because revision needs revise_from and the task ID never got out. And you lose fallback_scenes, the field that reports upstream quietly degrading shots - the render completes either way, so there's nothing left to warn you. Cuts you need to inspect go through the plain node; the concurrent version is for batches whose winners you'll eyeball.
Caching, cost, interruption
The seed is a cache key here, not a sampler setting - Motion Video never sends it to the API. fixed with unchanged inputs means ComfyUI reuses the cached future and you aren't re-billed; randomize means a fresh submission and a fresh charge. The receiver, by contrast, re-executes every run (that's deliberate), so queueing twice with a fixed seed gives you the same finished futures rather than a second batch of renders.
Interrupting cancels jobs that haven't started and signals running polls to stop, so bailing out of a half-finished batch doesn't keep the meter running invisibly.
Install
Nothing extra over the base pack. ComfyUI Manager, search ComfyUI Seedance, install, restart - or:
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
Dependencies are requests>=2.28 and that's it: no torch extension, no models in models/. You do need a key - a Seedance API Config node (https://api.seedance.nz plus api_key) wired into api_config, or SEEDANCE_API_KEY in the environment, or a config/.env in the pack folder.
Where it goes wrong
Queue with nothing connected and you get At least one concurrent Future is required. Wire an image future into a video receiver and it raises a type error naming both kinds - the two receivers aren't interchangeable. Under raise, one failing slot cancels everything still pending, including jobs already submitted and mid-render, so a single flaky render can strand work you've paid for; placeholder is the friendlier setting when each call costs real money. The usual pack-level irritations apply too: a missing key, SSL failures on Windows portable installs (refresh requests and certifi in the Python that runs ComfyUI), and expired result URLs - only local video files persist.
Inputs (19)
| Name | Type | Default | Description |
|---|---|---|---|
| model | COMBO | zhenzhen/motion-video | 2 options: zhenzhen/motion-video, zhenzhen/motion-video-pro |
| operation | COMBO | generate | 2 options: generate, revise |
| prompt | STRING | — | |
| duration | COMBO | 15 | 4 options: 15, 30, 45, 60 |
| ratio | COMBO | 16:9 | 3 options: 16:9, 9:16, 1:1 |
| resolution | COMBO | 720p | 2 options: 720p, 1080p |
| template_id | COMBO | api_default | 11 options: api_default, hype-kinetic, cyber-glitch, blueprint-tech, synthwave, swiss-minimal, +5 |
| title | STRING | — | |
| subtitle | STRING | — | |
| keywords | STRING | One keyword per line, or a JSON string array | 每行一个关键词,或 JSON 字符串数组 | |
| signature | STRING | — | |
| language | STRING | zh-CN | — |
| music | COMBO | auto | 2 options: api_default, auto |
| assets | STRING | Up to 4 persistent public HTTP(S) image URLs, one per line or JSON array; no local files | 最多4个持久图片直链,不接受本地文件 | |
| revise_from | STRING | Your completed Motion Video task ID | 自己已完成的动效视频任务ID | |
| revision_instruction | STRING | — | |
| api_configopt | SEEDANCE_CONFIG | — | |
| skip_erroropt | BOOLEAN | false | — |
| seedopt | INT | 00–18446744073709550000 | ComfyUI cache seed. Fixed reuses the cached result while all other inputs stay unchanged; randomize/increment/decrement starts a new execution. This value is not sent to models without documented seed support. | ComfyUI 缓存种子;Fixed 在其他输入不变时复用缓存,随机、递增或递减会触发新任务。未声明支持 seed 的模型不会收到此参数。 |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| future | SEEDANCE_VIDEO_FUTURE | — |