并发提交｜HappyHorse 1.1 视频生成

HappyHorse 1.1 is one of the more niche models in this pack, and honestly, that's exactly when you want the concurrent version. If you're already paying per clip on an API that isn't the headline Seedance model, you want to fire a few variants in one batch and pick the winner - not babysit sequential runs. This node is the concurrent-submit wrapper for the pack's HappyHorse 1.1 generator, covering the three models: happyhorse-1.1-t2v, happyhorse-1.1-i2v, and happyhorse-1.1-r2v (reference-image-to-video).

Mechanics are the pack standard: parameters get validated up front, then the job is pushed to a background video thread pool and you get a SEEDANCE_VIDEO_FUTURE handle back. Feed that into SeedanceConcurrent_Video_Await, and the receiver waits on up to ten futures in parallel, handing you finished VIDEO outputs in slot order. Three prompt variants of a t2v, or a t2v next to an r2v with character references - that's a one-receiver job.

The inputs that matter

• model - the three-way split above.
• prompt - required for t2v; optional for i2v and r2v. In r2v mode you can reference the input images directly in the prompt by writing 图1 / 图2 (the Chinese for "image 1/2") - that's how the upstream expects you to address reference material here, not @Image 1.
• seconds - 3 to 15, and note it does not support -1 for model-decided duration, unlike some siblings in the pack.
• resolution - 720p or 1080p only.
• ratio - forwarded as metadata.ratio; the API maps it to the upstream aspectRatio.
• first_image - required for i2v; for r2v it plays the role of 图1 (image 1), with reference_image2…reference_image9 filling out the reference set. Gaps in the middle get compacted to connected order, and the console tells you when it renumbers.

Plus the standard trio: api_config (or env var / config/.env), skip_error (placeholder video on failure instead of nuking the batch), and seed - a cache seed where fixed reuses cached output when nothing else changed.

How it runs

Same wrapper machinery as the rest: the i2v path sends first_image as images[0], r2v sends the compacted reference list, and the background worker handles upload → submit → poll → download with the MP4-header integrity check before a result counts as a real video. Because validation is synchronous, a bad combo (say, i2v with no first frame) fails at the node rather than halfway through a batch.

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; no weights. API key from the Seedance API Config node, SEEDANCE_API_KEY, or config/.env.

Where people trip

• i2v with no first_image - rejected up front. HappyHorse i2v only looks at the first frame, so don't hunt for a last-frame slot; there isn't one on this node.
• seconds = -1 - unsupported here. Pick a real number.
• Reference renumbering - skip reference_image4 and your reference_image5 becomes 图4. If your prompt says 图4 and the wrong face shows up, that's why.

The usual API-wrapper caveats apply: metered per clip, your prompt and frames leave the machine, and result links are temporary - save what you keep.