Developer Demonstrates Real-Time 1B Local AI World Model with Live Text Guidance

An independent developer trained a ~960M parameter pure transformer world model capable of transforming static images into interactive video streams on consumer GPUs. This updated model supports real-time keyboard action control (WASD) and dynamic text prompt switching mid-rollout, allowing users to modify the scene or character on the fly. Most generative video and world models require datacenter-class GPUs and suffer from high latency, making interactive local generation infeasible. Demonstrating real-time text-guided world modeling on consumer hardware like RTX GPUs and MacBooks opens up new possibilities for local game generation and real-time AI environments. The model features 28 blocks, 20 heads, block causal masking, and cross-attention for text, along with AdaLN terms for action inputs. Trained using diffusion forcing with independent frame noising, it runs at 50–60 FPS on an RTX 5090 (throttled to 12 FPS) by utilizing a sliding-window KV cache of 80 frames and running 2–5 diffusion steps per frame.