Dia2 2B

Dia2 2B is the second generation of Runware's hosted voice model, and "2B" is doing real work: roughly twice the parameters of Dia 1.6B, which buys you cleaner prosody and, more importantly, actual control knobs. Where the 1.6B node gives you one temperature, Dia2 splits the brain - separate temperatures for text tokens and audio tokens, plus top-K, plus a flag for how to treat reference audio. It's the TTS node you reach for when you care about the take, not just the words.

Like every node in this pack it runs in the cloud: no download, no local GPU, just a request over REST and an AUDIO tensor back. The catch is that with more knobs comes more ways to confuse yourself - so here's the mental model: text temperature is for how coherent the words come out, audio temperature is for how expressive the voice sounds, and 0 on either means greedy decoding (most consistent, least character).

What you set

• speech (required) - from the Runware Speech builder: text (with [Name] dialogue tags), voice (default Zephyr), speed, pitch, volume, language.
• settings.temperature - text-token sampling, default 0.6. Lower = more coherent.
• settings.audioTemperature - audio-token sampling, default 0.8. Lower = cleaner, higher = more expressive.
• settings.topK - limits selection to the K most likely tokens per step (default 50).
• settings.includePrefix - whether reference/prefix audio stays in the output or gets trimmed. Off by default.
• CFGScale - default 3 (1–30); the negative prompt only bites when guidance is on.
• outputFormat - MP3 / WAV / FLAC / OGG. advanced_json is your escape hatch for extra inputs.audios.

Output is audio (AUDIO).

How it works

The node builds a request with the runware:dia2@2b AIR and taskType: audioInference, ships it through the Runware SDK over REST, downloads the file, and decodes it with soundfile into a native AUDIO dict (falling back to torchaudio, then to a saved-file warning, if decode hiccups). The title bar prints the run cost. Because each request is a fresh cloud call, iterating on voice takes costs a few cents, not minutes of local queue time.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

The runaway-take failure mode is temperature creep - if outputs get weird, push audioTemperature down, not up. Voices sounding samey across lines? That's often the text temperature too low, or greedy decoding on. If you wired reference audio and it shows up in the output when you don't want it, that's includePrefix - flip it on to trim the prefix. And remember the guidance rule: negative prompts are dead weight until CFGScale goes above 1.