ProxyFace: Lightweight Local LLM Avatar with Real-Time Pixel Art Emotions

We are open-sourcing ProxyFace.com , a browser-based system that augments LLM interactions with real-time pixel art avatars expressing 8 distinct emotions. The entire pipeline runs locally without external API dependencies.

Technical approach The emotion classifier uses TinyBERT , chosen for its efficiency in browser environments while maintaining solid performance on short-sentence inputs.

Research artifacts

• Dataset : 3,200 labeled sentences for emotion classification

• Training : Google Colab notebook (free T4 tier)

• Assets : Model weights, dataset, and notebook

See screenshot:

We are sharing these resources to invite community contributions that improve both accuracy and inference speed. If you experiment with architectural changes, additional data, or optimization strategies, we welcome PRs and derived models.

Thank you for your attention!