并发提交｜Wan 2.7 Spicy 图生视频

"Spicy" in this corner of the ecosystem means one thing: the looser-filter edition. Wan 2.7 Spicy is the image-to-video flavor of Wan 2.7 that this pack's reseller gateway serves on its more permissive route, and this node is its concurrent submit wrapper - same behavior as the base Wan 2.7 Spicy 图生视频 node, but instead of rendering one clip inline it hands you a future to batch alongside other video futures. If you need a dozen variations of the same first frame and you know exactly what the base model's filter would have objected to, this is the node for the job.

The inputs that matter:

• first_image - required, sent as images[0]. The one frame the whole video springs from.
• prompt - up to 20,480 characters. On an i2v node this is optional in spirit - you can animate a frame with no prompt and let the model improvise, though you'll usually want some motion direction.
• seconds - 2 to 15. Wan 2.7 Spicy's range is tighter at the bottom than the Seedance nodes (those start at 4).
• resolution - 720p or 1080p, no 2k/4k tier here.
• negative_prompt - optional, forwarded to metadata.
• audio_url - optional public audio URL forwarded to metadata.audio_url, for sound you want synced into the clip. It must be a real http(s):// URL or the node refuses.
• prompt_extend - toggle to let the model expand your prompt, forwarded to metadata.prompt_extend.
• seed - -1 for random; non-negative values are actually forwarded to the API as metadata.seed. This is one of the few nodes in the pack where the seed does real work rather than just cache-busting.

Output is the single future (SEEDANCE_VIDEO_FUTURE), which goes into 并发接收视频（10 路） alongside up to nine other video futures.

How it works

Same concurrent architecture as every 并发提交｜ node here: settings validated up front, job handed to the video thread pool (10 workers by default, SEEDANCE_VIDEO_CONCURRENCY to adjust), handle returned immediately. The receiver waits on all slots, honors per-slot skip_error, and either raises on failure (failure_mode=raise) or drops a placeholder into the failed slot. The underlying pipeline is the pack's usual upload → submit → poll → download, just run N times concurrently.

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. Only dependency is requests; the key comes from api.seedance.nz.

Where people get burned

• Wan 2.7's reputation, honestly. The community's verdict on 2.7 was lukewarm-to-hostile - fine texture on close-ups, visible pose breaks on heavy motion. "Spicy" gets you past a filter, not past physics. Static, slow-pan, and character-consistency work is where it holds up; don't spend parallel credits on an action scene and be surprised.
• The audio_url validation. It must be a public URL, not a local file - there's no upload slot for audio on this node.
• Remembering the future isn't a video. Straight from submit node into the 10-路 receiver, then out.

Per-render cost and the data-leaves-the-machine caveat are the standing pack fine print. If you're batching spicy i2v, this node's whole reason to exist, it's probably worth setting a fixed seed per slot so you can actually compare what the prompt changes did.