Seed Audio 1.0

The most surprising node in this pack is an audio one. Seed Audio 1.0 is ByteDance's full-scene audio generator - the model that takes one prompt and synthesizes a complete soundscape: voice, environment, background music, sound effects, even dialogue - all from a single sentence. It's the audio equivalent of a text-to-image model, and it's the kind of thing that used to take a stack of libraries, samples, and a mixing session.

As with every ByteDance flagship, it's API-only - no weights, ever - which is the entire reason this node exists. It's a thin client for the hosted model: prompt goes in, an AUDIO clip comes out, no local runtime, no audio-plugins-in-ComfyUI madness.

What you set

Required: positivePrompt. The tooltip is the manual: describe the voice, environment, background music, sound effects, and dialogue you want in one prompt. "A rainy cafe in Paris, jazz playing softly, a woman ordering coffee in French" - that's a valid prompt for this model.

Then the extras that make it a production tool:

• speech - a RUNWARE_SPEECH socket for explicit text-to-speech (scripted narration) layered into the scene.
• audioSettings - a RUNWARE_AUDIOSETTINGS socket for format/sample-rate control.
• referenceImages - yes, images. Seed Audio can condition on visuals, which is how you get audio that matches a scene or character.
• advanced_json - the documented escape hatch for inputs.referenceAudios: reference clips in array order, which the prompt references as @Audio1, @Audio2, @Audio3. That's the audio-reference mode - feed it clips and the prompt can say "the guitar from @Audio1, the crowd from @Audio2."

Output is one audio (AUDIO), with outputFormat giving MP3/WAV/FLAC/OGG.

Where it fits

The obvious build is scoring a generated video: this pack's Veo/Wan/Grok nodes make the visuals, Seed Audio makes the soundtrack, and you mux them in your graph. It's also the one node that replaces an entire sound-design workflow for quick iteration - for foley and ambience on a concept clip, a single prompt beats an hour of hunting samples. The reference-audio mode (@Audio1 etc.) is the power feature worth learning: it turns the model from "invent a voice" into "recreate this specific voice/clip in a new scene."

Install and gotchas

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart, or install "Runware" from ComfyUI Manager; API key in Settings → Runware API key or RUNWARE_API_KEY. The pack pulls in soundfile specifically so audio decodes to a native AUDIO type without FFmpeg.

Gotchas: every run bills the account (title bar shows cost), and audio-ref mode is easy to get wrong - the @Audio1 references only work if you've actually wired reference audio through advanced_json, so a prompt full of @Audio1 tags with no references produces nothing useful. Read the tooltip before you chase that.