并发提交｜Hailuo 2.3 视频生成

Hailuo 2.3 is MiniMax's video API, and if you're using it from ComfyUI through this pack, this is the node you grab when "one clip at a time" stops being enough. It's the concurrent-submit wrapper around the pack's Hailuo 2.3 generator, so it keeps all six models (t2v standard/pro, i2v standard/pro, plus the two fast image-to-video paths) but fires them in the background instead of blocking the graph.

The deal works like every submit node in this pack: you set the parameters, it validates them, then kicks the job onto a background video thread pool and hands you a SEEDANCE_VIDEO_FUTURE. You connect that into the SeedanceConcurrent_Video_Await receiver (up to ten video futures per receiver), and it waits on the whole batch and returns the finished VIDEO values in slot order. So a storyboard pass where you want three t2v clips and a fast i2v render running simultaneously is just four of these nodes and one receiver.

The inputs that matter

• model - six choices: hailuo-2.3-t2v-standard / -pro, hailuo-2.3-i2v-standard / -pro, and hailuo-2.3-fast-i2v / -fast-pro-i2v.
• prompt - required for t2v, and capped at 2000 characters on Hailuo 2.3, which is tighter than most models in this pack. Keep it punchy.
• seconds - 6 or 10. Here's the trap: 1080p only supports 6 seconds. If you pick 1080p + 10s, expect a validation error before anything hits the API.
• resolution - 768p or 1080p.
• ratio - forwarded only for text-to-video; image-to-video follows your input image's aspect ratio, so the control disappears when it wouldn't do anything.
• first_image - required for the i2v/fast-i2v models. The README's field note spells out the constraint: the short edge must be greater than 300px and the aspect ratio must sit between 2:5 and 5:2. A skinny phone screenshot will get bounced.

The pack-wide trio is here too: api_config (or the SEEDANCE_API_KEY env var / config/.env), skip_error (placeholder video instead of killing the batch on failure), and seed - a ComfyUI cache seed where fixed reuses cached results so an unchanged re-run doesn't cost another paid call.

How it runs

Standard wrapper behavior: synchronous parameter validation up front (bad params die at the node, not in the middle of a batch), then upload → submit → poll → download in the background pool. Hailuo 2.3's i2v first frame goes in as images[0], and the download side does the whole H.264 ftyp-header check before it hands you a VIDEO.

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

Only dependency is requests, and there are no weights to fetch - the inference happens on the API side. Don't skip the API key step: Seedance API Config node, SEEDANCE_API_KEY env var, or config/.env, take your pick.

Where people trip

• 1080p + 10s - rejected; 1080p is 6s-only on Hailuo 2.3. Not a bug, it's the upstream spec.
• Tiny first frames - short edge under ~300px fails i2v. Upscale the frame before wiring it in.
• Placeholder videos - that's skip_error doing its job; the specific slot failed, and the receiver's status_json will tell you which and why (secrets scrubbed).

One thing to keep in your head: this is an API wrapper, so every clip is a metered cloud call and your prompt and reference image leave the machine. That's the whole category's trade - you get a closed model you can't run locally, and you pay per clip.