DreamForge AI Lab (Kazakhstan) Releases DF-World 0.1 Preview: Single-GPU Streaming World Model With Mid-Stream Reprompting and Dual First/Third-Person Control on Just 64 Hours of Adaptation Data

Self-funded Kazakhstan-based DreamForge adapts the LongLive/Wan video prior with a Matrix-Game-style residual action pathway to run a real-time controllable world model on a single consumer GPU. Uses only 64 hours of curated gameplay vs Genie's 30,000 hours, supports text/image/video initialization, mid-stream reprompting (e.g., redirecting a coastal rollout into a tsunami), and view-specific action LoRAs for first- and third-person modes. Acknowledged limit: no persistent spatial memory.