MiniMax H3 Context IR 提示词增强（3 合 1）

This is the one node in the pack that does not generate anything visual. MiniMax H3 Context IR is a prompt enhancer: you give it a rough video prompt, optionally some context (first/last frames, or a whole multimodal mood board of images, videos and audio), and it returns a richer, production-grade prompt via its result_text output. Then you wire that string into a video generation node's prompt input - the pack's own H3 OW video nodes are the natural match, but it'll feed any of the video models in this pack.

Three modes sit behind the dropdown: Text (pure prompt in, enhanced prompt out), Image (adds first/last frame context), and Multimodal (up to 9 images, 3 videos, 3 audios as context). It's the "tell me what to say to the model" step, and for anyone whose video prompts come out vague, it's quietly one of the most useful nodes here.

How it works

It submits to the same compat endpoint the video models use (POST /v1/video/generations), polls, and reads the enhancement from the response's result_text field. The node uses the standard seed-cache and skip_error controls, but it doesn't join the concurrent future system - it's a single-shot transformer, not a batch generator. Key real-world detail from the pack's testing: multimodal images go in the top-level images field, videos go into metadata.video_urls (plural), and audio into metadata.audio_url - and all three can ride in one task.

The inputs that matter

• model - minmax-h3-context-ir-text, -image, or -multimodal.
• prompt - the video prompt to enhance, 1 to 7000 characters.
• seconds - the target video duration, 4 to 15. It steers the enhancement's pacing advice.
• ratio - mode-dependent: Text requires one of the fixed ratios; Multimodal accepts adaptive or api_default; Image mode sends no ratio at all.
• image1/image2 (Image mode), image1..9 + video1..3 + audio1..3 (Multimodal) - the context pool.

Outputs: result_text (the enhanced prompt - the whole point), task_id, response.

Installing it

ComfyUI Manager → search "ComfyUI 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

Restart, key from api.seedance.nz/console, Seedance API Config into api_config.

Common issues

• It "doesn't work" - you're expecting a video. There is no video output. The result is text; connect result_text to a video node's prompt.
• Text mode rejects your ratio - Text requires one of the six fixed ratios; adaptive isn't allowed there.
• Image mode ignores your ratio - by design; it follows the frame.
• Multimodal references not landing - the pack verified the field mapping; if audio isn't landing, make sure it's an AUDIO input connected to audio1..3, and remember videos max at 3 and images at 9.

Use it the way the example workflows do: Context IR → enhanced prompt → H3 OW video. You get a noticeably more directed clip than prompting cold, and it costs you one extra API call.