LongCat AudioDiT Voice Clone TTS

What it is

If you've ever wanted a character's voice to carry across a whole project - same narrator, same dub, episode after episode - this is the node. LongCatVoiceCloneTTS is the reason to install this pack at all. Drop in a short clip of any voice, type what you want said, and the thing speaks in that voice. No training, no LoRA, no API. "Zero-shot" here isn't marketing fluff; the reference audio is conditioning, not a fine-tune.

How the cloning works

Mechanically it's elegant. LongCat encodes your reference clip through the same Wav-VAE it uses for generation, and those latents condition the diffusion transformer while it denoises. The model's paper claims it beats Seed-TTS on speaker similarity, and people testing it in the wild backed that up - the closest open clone quality most of us had heard up to that point. But it's a diffusion model, so it's sensitive. That's where the setup matters.

The inputs that matter

The inputs you care about:

• prompt_audio - your reference clip (AUDIO). 3–15 seconds is the sweet spot; the node warns if it's over 30.
• prompt_text - the transcript of that clip. The tooltip says it's "required" and improves quality significantly, and the code literally prepends it to what you're synthesizing, so don't skip it. If you do, you just get a console warning and worse output.
• text - what you want said in the cloned voice.
• guidance_method - defaults to apg here, and that's the right call; apg is recommended for cloning.
• steps - 16 default, 32 if you want the last bit of clarity.
• dtype - leave on auto. Here's the trap: fp16 silently kills voice cloning. The latent conditioning path overflows in fp16, you get NaNs cascading through the ODE solver, and the result is silence. The node notices and auto-upgrades fp16 to bf16 with a warning - but that only works on a GPU that supports bf16 (compute capability 8.0+). Older cards should pick bf16 explicitly or use fp32.
• seed - 0 is random. This matters more than people expect on a diffusion TTS; same prompt, different seed, noticeably different delivery.

Output is a single audio (AUDIO) noodle. PreviewAudio to hear it, SaveAudioMP3 to keep it.

Installing

Install is identical to the rest of the pack: ComfyUI Manager → search "LongCat AudioDiT" → install → restart, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-LongCat-AudioDIT-TTS
cd ComfyUI-LongCat-AudioDIT-TTS
pip install -r requirements.txt

The model auto-downloads on first run (bf16 recommended, fp8 if VRAM is tight), and remember to restart ComfyUI twice overall - the dependency auto-installer runs before the node registers on the first boot.

Where people get burned

Now the real-world part, because cloning is where people get burned. LongCat is extremely sensitive to reference audio. Loud or clipping input comes out as static - the README recommends normalizing to −3 to −6 dB peak, and community testers found even gentler levels (−20 to −25 LUFS) gave the cleanest clones. If your output buzzes or hisses, the fix is almost always the reference clip, not the node. Keep synthesized output in the 15–30 second range too; push past 60 seconds and words start repeating. And one note from the release threads: results vary by seed, so if a take sounds off, re-roll before you blame your reference.