wildminder/ComfyUI-VoxCPM

Implements a tokenizer-free diffusion-based TTS architecture built on MiniCPM-4 that models speech in continuous space rather than discrete tokens, enabling context-aware prosody generation. Includes native LoRA fine-tuning support within ComfyUI for custom voice style training, automatic model management with efficient VRAM offloading, and operates at 6.25Hz token rate for faster synthesis on consumer hardware. Integrates seamlessly with ComfyUI's node workflow system, supporting optional reference audio for voice cloning and compatible with multiple inference backends (CUDA, CPU, MPS, DirectML).

390 stars.