小珠光 LongCat

Voice cloning in ComfyUI means the LongCat-AudioDiT stack: a diffusion TTS model that takes a 3–15 second reference clip of a voice plus its transcript, and reads your text in that voice. It's genuinely good - and the original plugin has one maddening flaw. Even after you've downloaded the models, if HuggingFace is unreachable, the tokenizer check still tries to phone home and the whole thing hangs or dies. This node (小珠光 LongCat) is the pack's strict-offline fork of that voice-clone workflow: it reuses the LongCat modeling code but strips out every network call, so once your models are local, it runs forever without touching the internet.

Mechanically it's a careful fork. Where the original loader lists phantom "auto-download" model entries and calls snapshot_download on a miss, this node scans ComfyUI/models/audiodit/ and lists only what's actually on disk, throws a clear FileNotFoundError with the path and download instructions if nothing's there, and forces local_files_only=True on the transformers loads so nothing sneaks out to the network. It also has a 4-level fallback chain for the UMT5 tokenizer - your models folder, the sibling directory, your HF cache, then the original plugin's bundled copy - before it gives up. If you've had the LongCat plugin sit frozen on "loading tokenizer" for five minutes, this is the fix.

The inputs are the meat of any TTS node, and the tooltips do the explaining:

• text - what you want read aloud, in the cloned voice. Auto-splits long text into sentences.
• prompt_audio - the reference clip whose voice you're cloning; 3–15 seconds is the sweet spot.
• prompt_text - the transcript of that clip. Strongly recommended - without it, cloning quality drops noticeably.
• steps - ODE Euler steps, default 16 (4–64).
• guidance_strength - CFG/APG strength, default 4. Voice cloning is recommended at apg (see next).
• guidance_method - cfg or apg (default). The author recommends apg for cloning.
• device - auto / cuda / cpu / mps.
• dtype - auto / bf16 / fp16 / fp32; the tooltip warns fp16 numerically overflows, so bf16 is the clone recommendation.
• attention - auto / sdpa / sage_attention / flash_attention.
• seed, unload_mode (full_unload / offload_cpu / keep_gpu - the VRAM-management ladder), pause_between_segments (silence between auto-split sentence segments, 0.3s default), and volume_protect (optional, on by default: matches output RMS to the reference and guards against clipping).

One output: audio (AUDIO) - wire it to a save/combine node, and it pairs perfectly with the pack's video combine for muxing voice onto a clip.

The catch, and it's the important one: this is not self-contained. It deliberately reuses the ComfyUI-LongCat-AudioDIT-TTS plugin's modeling library, so you need that original plugin installed alongside, plus the LongCat model and UMT5 tokenizer in ComfyUI/models/audiodit/. The recommended path (from the README): install the original plugin once, let it download the models, then switch to this node and never look at the network again. librosa is in the pack's deps; registry install handles it.

Install the pack (ComfyUI Manager → ComfyUI-xiaozhuguang) and the LongCat plugin, drop the models in place, restart. If the dropdowns are empty, that's the strict-offline design telling you the files aren't where it expects - that's the feature, not a bug. For an offline voice-clone pipeline that just works, this is the one to reach for.