ITP Thesis Archive 2025

Project Description

This project is an AI companion creation platform designed to offer users emotionally meaningful, personalized interactions with digital characters. Users begin by uploading a portrait, writing a backstory, and crafting a first-encounter scene—building both the identity and emotional tone of their companion. The result is not just a character, but a nonjudgmental, always-available presence that lives in a private, user-defined space.
Once created, the AI companion engages in both text and voice conversations, responding in alignment with the personality and narrative the user designed. Real-time facial animations add expressive nuance, allowing the companion to visually react, deepening the emotional connection and sense of realism.
Beyond entertainment, the platform offers users a safe, intimate environment to explore love, comfort, or dramatic storytelling—without external judgment or limitations. Whether used for roleplay, emotional support, or creative expression, it gives users the power to shape how they are seen, heard, and cared for by their companion.

Technical Details

This project integrates a custom-trained graphics generation model using open-source resources like Hugging Face and Live Portrait for animated facial rendering. Voice interaction is powered by a LLaMA 70B-based TTS language model, specifically trained to specialize in memory and roleplay for emotionally consistent conversations. The frontend experience is built with React, using JavaScript, HTML, and CSS for a seamless, browser-based interface. I designed the UI/UX in Figma and handled full implementation.

Research/Context

This project began as a thesis exploration into AI assistants for daily planning and task tracking—driven by a personal struggle with procrastination. However, as the concept developed, I became more interested in the emotional presence of the assistant: how it looked, spoke, and offered support. I realized the value wasn’t in productivity metrics, but in the feeling of being understood and encouraged.
I researched tools like Replika, Character.AI, and Vtuber avatars, noting how visual personality and narrative design foster emotional connection. This inspired a shift toward building a customizable AI companion focused on roleplay, storytelling, and emotional intimacy.
To enable expressive interaction, I evaluated several technical tools. While options like DeepMotion and Live2D offered animation capabilities, I ultimately chose to train my own model using Hugging Face and Live Portrait, which gave me greater flexibility and control over visual expression. For voice, I implemented a LLaMA 70B-based TTS model, fine-tuned for long-term memory and immersive roleplay, allowing the companion to respond with emotional consistency. Compared to models like Chat-O1 or Fish, this solution is significantly more cost-effective and offers greater stability for sustained conversations.
This shift—from productivity to presence—redefined the project as a platform for co-creating deeply personal, emotionally intelligent AI companions.