What changed
AMD has introduced Micro-World 2, an interactive world model capable of real-time video generation. This model is designed to operate as a streaming system, continuously predicting future frames using only past observations and current control inputs. It achieves real-time performance (>30FPS) on a single AMD Instinct™ MI350X GPU, with the option to use a lightweight VAE for higher frame rates (42.49 FPS).
What this means in practice
Micro-World 2 extends the capabilities of its predecessor, Micro-World 1, by incorporating a unified Wan2.2 5B causal autoregressive model obtained through four-step distillation. It uses longer Capsaicin trajectories with frame-aligned action and camera-pose metadata, along with a ROCm-optimized inference stack, to sustain streaming generation for up to one minute. The model conditions on previously generated visual context, cached text features, and frame-aligned keyboard and mouse inputs to generate the next video chunk. It achieves a WBench average score of 76.68 and leads 8 of the 21 metrics reported in Table 4, with strengths in interaction, scene consistency, and action generalization.
Limitations and availability
The model is built on the AMD ROCm™ software ecosystem and optimized for AMD Instinct™ GPUs. It is fully open source, including the curated dataset, training pipeline, and inference code. The data collection process involves a frame-accurate Capsaicin dataset, with each clip containing synchronized keyboard and mouse inputs, text captions, and action annotations. The model is available for use with the specified hardware and software stack, though the supplied announcement does not specify exact deployment timelines or broader availability beyond the mentioned GPU and software ecosystem.
