并发提交｜FLUX 3 视频生成与草稿增强（8 合 1）

This is the batch version of the pack's FLUX 3 video node - the one that turns an 8-model generator (domestic/global text-to-video, image-to-video, video-to-video, and draft enhancement) into something you can run in parallel alongside other video jobs. If FLUX 3 is a model you actually reach for - the KB's ecosystem docs put the FLUX family at the center of the closed-model-vs-open-model conversation, and FLUX 3 Video is the video end of that line - then this is the node you use when one clip isn't enough.

It's a "concurrent submit" wrapper, so the mechanics are the same across the pack: instead of generating in place, it pushes the job to a background thread pool and hands you a SEEDANCE_VIDEO_FUTURE handle. That handle then gets wired into the SeedanceConcurrent_Video_Await receiver, which waits on up to ten of them in parallel and outputs the finished clips in slot order. Want three keyframe animations and a draft-enhanced scene in one run? Three of these nodes, one receiver, done.

The inputs that matter

The node mirrors the plain FLUX 3 node's parameter set, so most of it should look familiar:

• model - the 8-way dropdown: flux-3-video-t2v, i2v, v2v, draft-enhance, each with a global- overseas variant.
• prompt - required for T2V, I2V, and V2V. Draft Enhance doesn't use it; that model eats draft_cache instead.
• seconds (5–20), resolution (hd/fhd), ratio (8 documented options, auto default).
• draft (bool) - set this on a normal generation and the task returns a reusable draft_cache string. Wire that into a draft-enhance model's draft_cache input to refine the earlier result. That's the FLUX 3 headline feature and it chains across concurrent runs fine.
• audio_mode (api_default/enabled/disabled) and safety_tolerance (api_default or 0–4) - both stay untouched at the API default unless you need them.
• Optional media: image1…image10 for I2V keyframes (submitted in slot order), and input_video / video_url for V2V (public URL wins if both are set).

Plus the pack-wide trio: api_config, skip_error (returns a placeholder video on failure instead of killing the batch), and seed (ComfyUI cache seed - fixed reuses cached results, so a re-run that changes nothing won't spend another paid call).

How it runs

Mechanically it's one of the PURE_VIDEO_NODE_KEYS wrappers: it validates the model's parameters synchronously before the Future is created (so a bad param fails fast, not deep in the batch), then uploads your reference media, submits, polls, and downloads in the background thread pool. The video pool defaults to 10 workers, tunable via SEEDANCE_VIDEO_CONCURRENCY.

Installing

# ComfyUI Manager: search "ComfyUI Seedance" → install → restart.
# or:
comfy node install seedance
# 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

Dependency is just requests; nothing to download model-wise. API key comes from the Seedance API Config node, SEEDANCE_API_KEY, or config/.env.

Where people get burned

• Draft Enhance with no draft_cache - it's required; the node rejects it before submitting. Run a draft-enabled job first and pass its output through.
• Missing prompt on T2V/I2V/V2V - also fails fast at validation, before it touches the API, which is the good kind of failure.
• skip_error + placeholder videos - a placeholder in your output means that specific job failed; check the receiver's status_json rather than assuming the workflow is broken.

Like every node in this pack it's a cloud call - your prompt and keyframes go to the gateway server, results are metered per call, and direct links expire, so save what you love.