并发提交｜Wan 3.0 图生/参考生视频（8 合 1）

The Wan 3.0 concurrent submit node is the biggest single-dropdown in the pack: eight models - standard and "Prime" (the faster tier) flavors, domestic and overseas routes, and two task types each. It's the reference-to-video flagship, and it's where this pack shows off how many material slots it can stuff into one node. If you're batch-generating videos from reference material, this is the one.

The model dropdown decides everything about which inputs exist in spirit:

• wan-3.0-i2v / wan-3.0-global-i2v - standard i2v, domestic/overseas. image1 is the required first frame, image2 an optional last frame.
• wan-3.0-r2v / wan-3.0-global-r2v - standard reference-to-video, domestic/overseas. This is the big one: up to 10 images, 5 videos, 5 audios, plus optional file_url / link_url.
• wan-3.0-prime-i2v / wan-3.0-global-prime-i2v - Prime (high-speed) i2v.
• wan-3.0-prime-r2v / wan-3.0-global-prime-r2v - Prime r2v.

Shared controls: prompt (required for R2V, optional for I2V; up to 20,000 chars and you can reference Image 1 / Video 1 / Audio 1 in upload order), seconds (auto or 2–30), resolution (480P/720P/1080P - note the capital P here, this node uses the documented casing), ratio (six options, adaptive default), generate_audio, and seed (0 to 2,147,483,647, actually forwarded to the API).

Two fields deserve their own paragraph because they're the easiest to misuse:

• enable_thinking - standard overseas models only. Prime models don't support it and the node won't send it for them. The subtle bit: a standard Global R2V automatically enables it when you supply a file_url or link_url reference.
• file_url / link_url - R2V-only public document URL vs webpage URL, and they're mutually exclusive. Set both and the node refuses. These are for "here's a spec doc, make a video about it" workflows.

Output is the single future (SEEDANCE_VIDEO_FUTURE) into 并发接收视频（10 路）.

How it works

The usual concurrent architecture - settings preflighted synchronously, job handed to the video thread pool, future returned. Underneath, the reference materials upload (all of them, up to twenty slots of media if you're being ambitious), the payload is assembled per-model, and the task runs through submit → poll → download. skip_error isolates a failed slot; the receiver's failure_mode decides between raising and placeholding.

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 plus an api.seedance.nz key; no models to download.

Where people get burned

• The big reference mode is a lot of upload. Ten images, five videos, five audios is a heavy pre-generation step, and the gateway throttles per-token upload rates - the client waits and retries, so the node sits there for a while before the render even starts. Trim media, and know it's normal.
• file_url vs link_url exclusivity. Pick one. The validation is clear when you hit it, but it's also the field people randomly fill both into.
• enable_thinking doesn't apply to Prime. The node sends it only where the upstream documents it; the tooltip is your map.
• R2V requires a prompt. Unlike I2V, where an empty prompt is tolerated, R2V with no prompt gets rejected.

Wan 3.0 itself is API-only and, like 2.7, has no local weights - this node is the way you use it from ComfyUI at all, and the Prime variants are the pack's answer to "I don't want to wait forever per render."