MiMo 文本音色设计

Here's the node from this pack that has no local equivalent. Preset TTS gives you a fixed cast of voices; voice cloning needs a recording. MiMoVoiceDesignTTS takes neither - you write a sentence or two describing the voice you want ("a warm young woman, clear Mandarin, medium pace") and Xiaomi's MiMo-V2.5-TTS-VoiceDesign builds it on the server and speaks your text. That "describe a voice from scratch" trick is exactly what hosted TTS does well and open models mostly can't: local tools either clone existing voices or cycle through presets, and designing an imaginary character's voice on the fly is the newest trick in the space, the same capability the community benchmarks against Mimo and Qwen3-TTS-VD.

It's an API wrapper, same shape as the rest of the pack. The node POSTs your voice_description as a user message, your text as the assistant turn, and the WAV comes back decoded into a standard ComfyUI AUDIO output (waveform + sample_rate). Nothing runs locally - no GPU, no model download, no extra pip packages. You pay per call and your text and description leave the machine; that's the deal with every hosted TTS, and it's the reason to reach for this only when a local model can't do the job. For "I need a specific fictional character's voice and I have no reference audio," that's exactly the case.

Inputs worth knowing:

• voice_description - the 1–4 sentence voice spec. The default is a Chinese example, so keep it in the same language as your text for best results. This is the field you'll iterate on; the whole quality of the output lives here.
• text - what gets spoken.
• optimize_text_preview - a boolean. When on, the server may polish the target text before speaking, which can smooth out awkward phrasing at the cost of speaking something slightly different from what you typed. Default off; flip it on when the delivery feels stilted.

Plus the shared api_key (or the MIMO_API_KEY env var - recommended, since a key in the node field gets baked into your workflow JSON and will leak if you share it), timeout (default 180s) and max_retries (default 2, handles 429/rate-limits and temporary 5xx errors). Output is one audio node, wired to preview, Save Audio, or a lip-sync pipeline if you're making a talking head.

Install is the same as its siblings:

cd ComfyUI/custom_nodes
git clone https://github.com/TimekeeperXY/ComfyUI-MiMo-TTS.git

Then restart. ComfyUI Manager will also find it if you search "MiMo TTS". No extra dependencies - the pack only uses av and torch, which ComfyUI already ships.

Two honest caveats. First, this is a paid API, and the bill is where you'll feel it if you brute-force your way to a voice - describe carefully, then pay once. Second, because the same inputs hit ComfyUI's execution cache, a "no-op re-run" returns the cached audio and doesn't re-charge; change any input to force a fresh synthesis. And if the designed voice misses the mark, rewrite the description rather than adding a bunch of style appendices to your text - the description is the entire voice model, and the text is just the lines.