ZONOS2 Voice Generation

This is the node to reach for when you just want to hear what ZONOS2 sounds like. Wire the zonos2_model output from the ZONOS2 Model Loader into it, type some text, and you get a mono ComfyUI AUDIO at 44.1 kHz out the other end. No API, no key, no voice reference required - the whole thing runs on your GPU.

It's the plain-text branch of the pack. If you want the party trick, that's the ZONOS2 Voice Clone node. But the generation node is where you sanity-check that the model loaded right and the settings you're about to reuse actually behave.

How it works

Under the hood this is a real autoregressive language model, just over audio instead of words. Your text gets tokenized as UTF-8 bytes, and a sparse MoE backbone generates DAC code frames - discrete audio tokens at 44.1 kHz - which a Descript DAC decoder turns back into a waveform. That's why the sampling knobs look like they were lifted off an image sampler: temperature, top_k, top_p, min_p, plus a repetition penalty tuned for audio (ZONOS2's stock sampling pipeline uses Min-P at 0.18 by default, with top_p disabled - leave it that way).

The model also accepts a pile of "quality conditioning" buckets - speaking_rate, loudness_lufs, estimated_snr, maximum_pause, estimated_bandlimit_hz, leading_silence, trailing_silence. Each is a dropdown of ranges (bytes-per-second, LUFS, dB, seconds) and every one defaults to default, which means "don't condition on this." Real talk: a beginner should ignore all seven. They're for matching the model's output to a recording's characteristics, not for day-to-day use.

The inputs that actually matter

• text - what to say. Anything not empty works; the integration uses ZONOS2's raw UTF-8 path, so written numbers and dates won't always be normalized the way you'd want - "2026" may read as digits. This is a model-level quirk, not a bug.
• max_new_tokens (default 1024, up to 6000) - the ceiling on how many audio frames get generated. Output stops early when the model emits end-of-audio, so if your clip keeps cutting off mid-sentence, raise this.
• temperature (default 1.15), top_k (106), min_p (0.18) - the randomness stack. Lower temperature for steadier delivery, around 0.8–1.0 if a voice is drifting run to run.
• seed - 0 means random; set a positive value to make a run repeatable. Worth doing once you find a delivery you like.

The single output is audio, a native ComfyUI AUDIO at 44.1 kHz. It feeds straight into ComfyUI's Save Audio / SaveAudioMP3 nodes, or into any audio-consuming node in your video workflow.

Installing and the first run

You'll need the whole pack: ComfyUI Manager → search ZONOS2 TTS, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/Zonos2_TTS-ComfyUI.git
../venv/bin/python Zonos2_TTS-ComfyUI/install.py

Restart, then let the loader pull the model. Watch the Transformers version - the pack needs 5.0–5.12 (baseline 5.3.0) and won't run on 4.x, and install.py won't upgrade it for you. And be patient with the first generation: loading the ~14 GiB BF16 checkpoint plus the DAC takes a while, and it's normal for the first load to look frozen.

Troubleshooting

• Output stops too early → raise max_new_tokens; the model can legitimately end early on its own.
• Sounds wobbly or unstable run to run → try temperature around 0.8–1.0 and set a fixed seed.
• Numbers and dates read wrong → the raw UTF-8 path doesn't do full NeMo text normalization; spell out tricky values or run text through a normalizer upstream.
• CUDA OOM → the loader's FP8 preset is the memory-saver; also unload other models and lower max_new_tokens.

One honest caveat: with no voice reference, this node runs without any speaker embedding at all, so you get ZONOS2's bare, un-conditioned voice. It's great for testing and for narrator-agnostic audio. The moment you want a specific voice, that's the clone node's job.