VoxCPM2, CosyVoice and Qwen3 TTS: ComfyUI setup

This Russian-language tutorial compares local AI voice workflows for consistent character speech in LTX videos. The presenter tests VoxCPM2 in ComfyUI, a separate VoxCPM2 Portable build shared by a third party, CosyVoice and Qwen3 TTS with Russian dialogue.

The VoxCPM2 setup covers installing custom nodes through ComfyUI Manager and their pip dependencies. After an incomplete model download, the presenter manually adds a missing 360 MB audio VAE file. He demonstrates voice design and cloning, then writes numbers as words to correct pronunciation. The cited 12-24 GB VRAM requirement concerns LoRA training, which he does not demonstrate.

VoxCPM2 Portable asks users to select a GPU configuration. His RTX 3060 setup runs without Flash Attention 2. He reports poorer results at its initial settings than in ComfyUI, with improvements after adjusting diffusion steps and generation retries.

CosyVoice replaces the voice in a recorded dialogue. The presenter says it preserves pauses and pronunciation quirks; the demonstrated workflow limits audio to 30 seconds.

For Qwen3 TTS, he discusses 0.6B and 1.7B models, with voice design limited to 1.7B in this workflow. He reports that 1.7B runs well with 12 GB VRAM, but encounters dependency conflicts in an existing build. Voice cloning requires text matching the reference audio; he warns that missing or inaccurate text can cause errors or stalls.

The creator links the example workflows and prepared portable builds through Boosty.