并发提交｜MiniMax H3 OW Fast 视频生成（5 合 1）

The "Fast" tier of MiniMax H3 OW is the pack's budget lane into the H3 family - cheaper and quicker than the full H3, at 480p/720p. The KB's closed-source doc notes H3's local weights come with a license that geofences the US/EU/UK/Korea; the API you're calling here sidesteps that whole question (it's a service, not a file you hold). This node is the concurrent-submit version of the pack's 5-in-1 Fast node, and the interesting part is the two audio-driven models - give the model one image and one audio clip, and it animates the image to the audio. That's genuinely unusual in this pack.

Standard wrapper mechanics: params validated up front, job pushed to a background video thread pool, SEEDANCE_VIDEO_FUTURE returned. That goes into SeedanceConcurrent_Video_Await, which waits on up to ten video futures and returns finished VIDEO values in slot order.

The inputs that matter

• model - five choices:
  • minimax-h3-ow-i2v-fast - first-frame image-to-video; image1 required, prompt optional.
  • minimax-h3-ow-r2v-fast - reference video from 1–9 images; prompt required.
  • minimax-h3-ow-fl2va-audio-drive-fast and ...-ref2va-audio-drive-fast - audio-driven: exactly one image1 plus one audio, prompt optional.
  • minimax-h3-ow-t2v-fast - pure text-to-video; prompt required, no images.
• seconds - 5, 10, or 15.
• resolution - 480p or 720p.
• ratio - the full eight: 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9.
• image1…image9 - I2V uses only image1; R2V opens the whole bank.
• audio - appears only for the two Audio Drive models and is required there; exactly one local audio gets uploaded and sent as a single-element metadata.audio_urls array.

Plus api_config, skip_error (placeholder video on failure), and the cache seed (fixed reuses results when nothing changed).

How it runs

The front-end dynamically shows 0, 1, or 9 image slots depending on the model - pick t2v-fast and the image slots vanish; pick an audio-drive model and a single audio input appears. Local audio is converted to WAV and uploaded before submit. The background worker does the poll-and-download, and an H.264 MP4 check stands between "returned" and "success."

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; no weights. API key via Seedance API Config, SEEDANCE_API_KEY, or config/.env.

Where people trip

• Audio Drive models are strict - one image, one audio, no more, no less. The node enforces it, so don't hunt for a second audio slot.
• t2v-fast with images attached - ignored, not an error, but pointless. The model uses prompt only.
• Missing prompt on t2v/r2v-fast - rejected at validation, before the API call. The fast failure is the good kind.

The usual pack caveats: metered per clip, your image and audio go to the gateway, and result links are temporary. The audio-driven models are worth one honest look - that's the feature people actually want out of H3.