并发提交｜Hailuo H3 视频生成

Hailuo is MiniMax's video model family, and H3 is the one that made waves when it dropped - the KB's closed-source-models doc frames it as the first credible open-weights answer to Veo's native-audio gap (and notes the license carries a weird geographic restriction for the local weights; the API you're calling here is a different door entirely). This node is the concurrent-submit wrapper around the pack's H3 generator: six models - domestic and global- overseas versions of t2v, i2v, and the big multimodal one - that fire in the background instead of blocking your graph.

Same wrapper pattern as every submit node in the pack: you fill in the params, it validates them synchronously, then pushes the job to a video thread pool and hands you a SEEDANCE_VIDEO_FUTURE. Wire that into the SeedanceConcurrent_Video_Await receiver and it'll wait on up to ten futures in parallel, returning finished VIDEO values in slot order. If you're comparing a t2v take against a multimodal take, this is how you run them at the same time instead of back-to-back.

The inputs that matter

• model - hailuo-h3-t2v, hailuo-h3-i2v, hailuo-h3-multi, plus the three -global- variants.
• prompt - required for T2V and Multi; optional for I2V (you can drive an i2v mostly off the frame). In Multi mode you address your reference material directly in the prompt with @Image 1, @Video 1, @Audio 1.
• seconds - 5 to 15.
• resolution - 768P or 2K. That's it; no 1080p middle ground.
• ratio - used by T2V and Multi (including adaptive); I2V ignores it and follows the input frame.
• image1…image9 - I2V uses image1 as the required first frame and image2 as an optional last frame; Multi opens the full nine-image bank plus video1…video3 and audio1…audio3 as references.

Pack-standard extras: api_config, skip_error (placeholder video instead of killing a batch), and seed as a ComfyUI cache seed - fixed means an unchanged re-run reuses cached output rather than spending another metered call.

How it runs

Classic wrapper mechanics: the Multi path uploads your local images, videos, and audio automatically and maps them to the right reference fields in the request; the receiver does the polling and download, and a clip only counts as success after it passes the MP4 header check. The batch layout means a failed slot doesn't take the others down - raise mode will stop and name the slot, placeholder mode keeps the survivors.

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

Just requests for a dependency, zero model downloads. API key via Seedance API Config, SEEDANCE_API_KEY, or config/.env.

Where people trip

• Multi with no reference material - the model needs at least one connected image/video/audio; a pure-prompt Multi job will fail validation. That's what the t2v model is for.
• @Image 1 refs that don't resolve - the tag numbering is by connected slot order, and the pack compacts gaps (skip image2, and what you connected as image3 becomes @Image 2). The console prints a hint when it renumbers.
• I2V ratio confusion - don't fight the input frame; I2V doesn't send a ratio at all.

It's a cloud call end to end - prompts and reference media go to the gateway, and each clip is metered. Save results you like promptly; the result links are temporary.