Hex AI Girlfriend: Windows setup with Qwen3 and ComfyUI

This Chinese-language tutorial builds Hex AI Girlfriend, a browser-based voice companion. CosyVoice2 provides speech synthesis in the linked project. The presenter explains how to replace its reference recording and matching transcript to customize the voice, recommends a short recording, and demonstrates a talking avatar with conversation logs. The wrapper is distinct from the underlying models.

The Windows walkthrough installs and configures WSL2, copies the project and model bundle, runs installation/start scripts and checks the services. A local Qwen3-8B Q4 model supplies replies through llama.cpp on port 8200. The presenter discusses context length, smaller models and disabling thinking to shorten response delay.

ComfyUI Desktop uses a Wan2.2 workflow to turn an image into a short avatar video. The presenter corrects his earlier explanation: the still-image prompt goes to GPT, while the video prompt goes to ComfyUI. He recommends a loop with matching first and last frames and reports slow generation on his RTX 4060 Ti. The project later applies a lightweight lip-sync model, whose mouth detail can blur.

The remaining setup covers microphone/speaker selection, avatar IDs, persona instructions, Windows reserved ports and logs. The presenter describes local runtime after downloading dependencies and slower lip sync on CPU; this is not an audited privacy guarantee. English speech and mobile access are possible future work, not features demonstrated here.