Seed Audio 1.0 · Text to Audio

This is not a text-to-speech node, and treating it like one is how people end up disappointed. Seed Audio 1.0 takes a description of an entire audio scene - two people talking over each other, a door, rain, a radio in the corner - and renders all of it as one mixed piece of audio. Dialogue, ambience, music and effects, in the relationship you described.

That's a different job from Seed Speech 2.0 in the same pack, which does one voice reading one script. If you need narration, go there. If you need the sound of a scene, this is the one, and at 54 credits per run the price tells you it's thinking bigger than a sentence.

How it works

One node, one model: bytedance/seed-audio-1.0/text-to-audio, called on PopNow's servers. The prompt goes over as a scene description; the result comes back as a ComfyUI AUDIO value, the same dict of waveform plus sample rate that ComfyUI's core audio nodes speak, so Save Audio writes it out and anything downstream that takes AUDIO just works.

Because it's a hosted call, the usual terms apply: metered, quoted before the run, refunded if it fails, and your prompt leaves the machine. There are no local audio models involved and no VRAM cost, which for a model that mixes multiple voices is the honest reason to use it - the local audio stack in ComfyUI is thin, and everyone doing this locally is chaining several models together to approximate one pass of this.

The inputs that matter

• prompt - the whole scene: who speaks, how, what they say (in quotes), plus music, ambience and effects. Write it like a script mixed with a sound-design note. Speaker-by-speaker paragraphs work better than one run-on sentence.
• reference_audio_urls - up to three clips, addressed as @Audio1 to @Audio3 in your prompt. This is the interesting control: give it two voice samples and write "@Audio1 says ..." and the scene uses those voices. Note the tooltip's constraint - not with image_url.
• image_url - an optional IMAGE to build the audio around; the scene is generated from what's depicted. Also mutually exclusive with the audio references, so pick a lane.
• speed, pitch, volume - 0.5–2, −12 to +12 semitones, 0.5–2. Blunt global knobs, not per-character. Fine for tuning the read; don't expect to audition one voice at a time.
• output_format - mp3 (default), wav, opus, or pcm. Take wav if you're going to edit or process it; pcm is headerless 16-bit little-endian stereo, which is only what you want if something downstream expects exactly that.
• sample_rate - defaults to "model default" (24 kHz). Set 48000 if you're dropping it into a video timeline, 44100 for music-shaped work.

Outputs

audio - and it can be a batch: if the model returns several takes, they arrive as a batch, and Save Audio writes one file per item. Check your output folder before assuming you got one file. The second output, task id, is your handle for the API console and support.

Install

Search PopNow in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Add a key from popnow.ai/api/console/keys in Settings > PopNow or via POPNOW_API_KEY, restart again, done. The pack has no pip dependencies at all - if you've been burned installing audio node packs that drag in a dozen packages and a Torch downgrade, this is a pleasant change.

Common issues

• "PopNow 400 naming a field" - almost always the reference-image / reference-audio mutual exclusion. Use one or the other.
• PopNow 40201 - credits. This node is billed per run, not per second, so a run costs the same whether the scene is two seconds or forty.
• Output sounds like a monologue - the prompt named one speaker too well. Name two, give them different tones, and let them overlap.
• @Audio1 ignored - reference clips need to actually be connected as an autogrow socket, and there's a limit of three. More than that doesn't stack.
• Cancelling stops your wait, not the run - it finishes on PopNow and bills if it succeeds.

Pair it with a Seedance clip that has generate_audio off, and you've got a two-pass pipeline: picture from one node, the sound of the scene from another. Which is closer to how anything gets made.