This repository provides a simple and practical workflow for building a local Live2D-based AI desktop robot. If you want to quickly try it out, just clone this repository, install the dependency and run main_live2D.py.
| Features | Status |
|---|---|
| Live2D Character | ✅ |
| Speech recognition (Google API) | ✅ |
| Test-to-Speech (Online Edge-tts) | ✅ |
| AI communication(Qwen model) | ✅ |
| Streaming text generation | ✅ |
| Personality Setting (System and Dynamic prompt) | ✅ |
| Tools like weather searching | 🚧 |
| Short-term Memory (Replay) | 🚧 |
| Long-term Memory (RAG) | 🚧 |
| Emotion system | 🚧 |
| Interaction of mouse or user's action | 🚧 |
Notes
- ✅ means a feature that is currently available and working properly.
- 🚧 indicates a feature that is planned, under development, or requires further optimization.
This subject is supported by open weights model. Even if we implement different prompt and add the memory, the answer is still stiff meaning differ from the normal speech. Perhaps the better model or having training could relieve this.
This project is inspired by Neuro-sama and AkagawaTsurunaki's video on bilibili.
This project is done by vibe coding(cursor).
This project makes use of the following open-source projects and resources:
-
live2d-py (Python bindings for Live2D Cubism SDK)
https://github.com/Arkueid/live2d-py the main repository: https://github.com/AkagawaTsurunaki/ZerolanLiveRobot -
Live2D Cubism SDK (Official SDK)
https://www.live2d.com -
Qwen2 7B (4-bit quantized model, Hugging Face)
https://huggingface.co/Qwen
Thanks to the authors and communities for their excellent work.
