Inworld TTS-1.5 Mini

Narration is a draft most of the time. You don't want the best voice model in the world for the fifth revision of a script you're still cutting; you want something fast and cheap that proves the timing works. Inworld TTS-1.5 Mini is exactly that - the lighter sibling of the 1.5 Max voice node, and structurally identical to it.

It's a hosted TTS model served through Runware, so it lives in your ComfyUI graph as a proper node with an AUDIO output rather than a hand-rolled API call. The pitch: use Mini to iterate on scripts and timing, swap in Max (or the realtime TTS-2) for the final render. Same family, same voices, lower bill.

How it works

The required input is a speech socket of type RUNWARE_SPEECH - the text and voice live upstream in a Runware Speech builder, not on this node. In that builder you set your text (with [Sam] Hello [Bob] Hi there style tags for dialogue, plus markup like [laughs] or [short pause]), pick a voice from the named dropdown (Zephyr by default), and optionally set speed, volume, and language.

On the model node itself: settings.temperature (0.1–2, default 1.1) tunes expressiveness - lower for flat, deterministic reads, higher for animated delivery. settings.textNormalization expands numbers and abbreviations ("Dr." → "doctor") before generation, which you want for prose and not for product codes. numberResults tops out at 4, and outputFormat is MP3/WAV/FLAC/OGG. The Runware Audio Settings builder wires into audioSettings if you need specific channels, bitrate, or sample rate.

Output is a ComfyUI AUDIO value (waveform + sample rate), decoded locally with soundfile - straight into a Save Audio node, exactly like the pack's audio workflow.

Installing it

Same official pack:

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

Restart, then set your API key in ComfyUI Settings → Runware API key or via RUNWARE_API_KEY. Requirements include soundfile for audio decode - nothing heavier.

Where people get burned

The schema is a copy-paste of Max, so the only real decision is which node you're using - and people forget Mini is the budget lane, then wonder why a hero VO track sounds less polished than the Max render next to it. That's not a bug; that's the tiering. And the pack-standard reminders: the text lives in the Runware Speech builder, not here, and this is a metered API - Mini is cheap, but numberResults of 4 across twenty script versions is still charges accumulating.