VibeVoice Loader

VibeVoice is Microsoft's diffusion TTS/voice-clone model, and it's become one of the most-liked TTS options in the ComfyUI world - there's a dedicated wrapper with hundreds of stars out there. AIIA's VibeVoice Loader is just how this pack pulls that model in. It has two dropdowns and one output, which is exactly as boring as a loader should be. The interesting decisions happen before you touch it.

What the two inputs mean

• model_version - 0.5B (Realtime), 1.5B (Standard), or 7B (Standard). This is the real fork in the road:
  • 0.5B Realtime - the streaming model. Fastest, lowest VRAM, most languages (English, Japanese, Korean…), but it has no zero-shot cloning and requires a voice preset from Microsoft's .pt library. It also needs an 8K context, so it's the low-latency dialogue pick, not the audiobook pick.
  • 1.5B Standard - the sweet spot. ~3GB VRAM, 64K context, proper zero-shot cloning. The README's testing found 1.5B nearly indistinguishable from 7B in style and content at double the speed and a quarter of the memory. This is the one I'd reach for.
  • 7B Standard - ~14GB VRAM, 32K context, highest quality ceiling, and slow. Use it for offline, no-budget-limit generation where the extra expressiveness earns its keep.
• dtype - fp16 / bf16 / fp32. fp16 is the default and fine for most cards; bf16 is worth it on Ampere+ hardware.

The output is a vibevoice_model object. That feeds the pack's VibeVoice TTS nodes (Standard or Realtime) - the loader itself never makes a sound.

The part that bites: the model files

The loader auto-downloads from HuggingFace, but it looks in a specific place: ComfyUI/models/vibevoice/. The README's documented layout is models/vibevoice/microsoft/VibeVoice-1.5B/ (7B lives under a vibevoice/ subfolder, vibevoice/VibeVoice-7B/). VibeVoice is built on the Qwen2.5 tokenizer, and if the download skips the tokenizer files the node explodes with a tokenizer error. The manual fix is fetching tokenizer.json, tokenizer_config.json, vocab.json, and merges.txt from the matching Qwen2.5 model and dropping them in the model folder.

Two more gotchas worth knowing:

• The pack ships its own fixed vibevoice_core - don't drop modeling_vibevoice_*.py scripts into the model dir or you'll get conflicts.
• The speed control on the TTS side uses the system sox command for time-stretching. On Linux that's sudo apt-get install -y libsox-dev sox; on Windows you have to install the binary and add it to PATH, or speed will silently fail.

Install

Install the pack through ComfyUI Manager (search ComfyUI_AIIA) or git clone https://github.com/havvk/ComfyUI_AIIA.git into ComfyUI/custom_nodes/ and restart. First run grabs the model - the 7B is a chunky download, so start with 1.5B unless you've got a reason not to.