Developer Zonghaoyuan launched InfiPlot on GitHub on June 2, 2026, as the world's first interactive story game that generates all text and images in real-time using AI. The project gained 190 GitHub stars within three days by combining multi-agent coordination with predictive generation to create seamless narrative experiences.
Multi-Agent Framework Coordinates Five Specialized AI Roles
InfiPlot employs a multi-agent framework where five AI roles work in concert to generate dynamic content. The Architect manages overall story structure and pacing, while the Screenwriter handles dialogue and narrative progression. A Character Designer maintains consistency across scenes, the Scene Director establishes environmental context, and the Painter coordinates visual generation.
The technical stack uses DeepSeek's deepseek-v4-flash for text generation, Runware's FLUX.2 klein 9B for images, Google's gemini-3.5-flash for vision analysis, and Xiaomi MiMo's mimo-v2.5-tts for text-to-speech. The frontend runs on Next.js with TypeScript, deployable to Vercel or Cloudflare Workers.
Predictive Generation Eliminates Perceptible Loading Times
Unlike traditional interactive fiction with predetermined assets, InfiPlot generates every element dynamically based on user choices. The system pre-generates likely next scenes during user reading time, creating the illusion of instant transitions between narrative branches. Stories consist of sequential scenes combining AI-generated backgrounds with dialogue trees.
The project optimizes image generation costs by only creating new visuals when choices lead to genuinely new locations or times. For dialogue-heavy scenes that remain in the same setting, the system reuses existing backgrounds while generating new text content. This selective approach balances visual appeal with API costs while maintaining narrative immersion.
Flipbook-Style UX Demonstrates Practical Multi-Agent Coordination
InfiPlot's user experience resembles interactive fiction but with real-time visual generation at each decision point. The multi-agent system maintains character and story coherence across scenes by coordinating context between specialized roles. Each agent operates with awareness of decisions made by other agents, ensuring consistent characterization and plot development.
The architecture solves several technical challenges:
- Latency masking through predictive generation hides model inference time
- Context-aware image generation produces visuals that match narrative state
- Multi-modal coordination synchronizes text, image, vision, and audio models
- Cost optimization through selective asset generation
Open-Source Release Enables Infinite Narrative Possibilities
The project released under AGPL-3.0 license with Chinese and English documentation. Unlike traditional interactive fiction engines like Twine or ChoiceScript with pre-written branches and static assets, InfiPlot enables infinite narrative combinations through real-time generation. Characters can evolve dynamically, and storylines adapt to user choices without predetermined limits.
The 190 stars gained in three days suggests organic developer interest, particularly from the Chinese market where interactive AI content is gaining traction. Potential applications extend beyond gaming to interactive training simulations, educational storytelling, therapeutic narrative exercises, and procedural content generation research.
Key Takeaways
- InfiPlot uses a five-agent system combining DeepSeek V4, FLUX.2, Gemini 3.5, and MiMo for real-time story and image generation
- Predictive generation pre-creates likely next scenes during reading time, eliminating perceptible delays between narrative branches
- The system optimizes costs by only generating new images when the story genuinely changes location or time period
- The project gained 190 GitHub stars within three days of its June 2, 2026 launch
- Released under AGPL-3.0 license, the TypeScript-based project deploys to Vercel or Cloudflare Workers with free playable demo