Introduction

AMD just open-sourced Micro-World 2, a world model you can drive with a keyboard and mouse in real time. The AMD AIG team published the weights at amd/Micro-World-2, the code in the Micro-World GitHub repo, and a technical blog dated October 21. You give it a first frame (or just a prompt) and press keys, and it streams the next frames of a first-person video as you go. AMD says it runs at more than 30 FPS on one AMD Instinct MI350X GPU and keeps going for up to a minute.

Micro-World 1 was built for short clips made offline. Version 2 is built to stream, which is the part that matters if you want something that feels like a game instead of a video render.

What shipped

  • Name and team: Micro-World 2, from AMD's AIG team (Yu Geng, Wensong Chan, He Cui, Dong Zhou, Dong Li and Emad Barsoum)
  • Modality: action-controlled video world model. It takes text or image to video, steered by keyboard and mouse input frame by frame
  • Size and base: 5B parameters, built on Wan2.2-TI2V-5B. The repo ships a causal init, a causal LoRA, an action module, a 4-step DMD student LoRA, and an optional lightweight decoder. The Wan2.2 base weights come separately from Wan-AI
  • Output: 352×640 video, 4 denoising steps per chunk, no CFG
  • Speed: 32.12 decoded FPS with the original Wan2.2 VAE and 42.49 FPS with the lightweight decoder, both on a single MI350X
  • Open vs closed: open weights (OpenRAIL license on the model card), inference and training code, plus the action-labelled Capsaicin dataset
  • Where to run it today: locally, through the ROCm Docker image or a conda setup from the repo. There's a browser demo you serve yourself, but no hosted demo or ComfyUI node yet

How it works

Wan2.2 normally denoises a whole clip at once, with every frame looking at every other frame. That can't respond to a key press in the middle of a clip. AMD swapped in block-causal attention over small chunks of latent frames, with a rolling KV cache, so each chunk only depends on what came before plus your current input. The VAE also decodes one latent frame at a time, so clip length is no longer capped by memory.

To get speed, the team distilled the model to 4 steps with DMD. The teacher and critic in that distillation also see the actions, which AMD says is what keeps the fast student responsive to your input.

The training data is the interesting part for game folks. AMD recorded it in Capsaicin, its own open-source real-time renderer. It loaded open 3D scenes, placed cameras, ran scripted keyboard and mouse controls, and saved the video with per-frame action labels. Open-domain Mixkit clips supply the visual variety on top of that. All training took about four days on eight MI350X GPUs.

On WBench, AMD reports an average of 76.68, just ahead of Sana-WM at 76.00, and first place on 8 of 21 metrics against seven other action and world models. Those are AMD's own numbers, not an independent benchmark.

Try it

git clone https://github.com/AMD-AGI/Micro-World.git && cd Micro-World
docker build -t microworld:latest .   # ROCm image, recommended

hf download amd/Micro-World-2 --local-dir ckpts
hf download Wan-AI/Wan2.2-TI2V-5B \
  --local-dir ckpts/models/Diffusion_Transformer/Wan2.2-TI2V-5B

# fixed action presets (w / s / a / d / mouse probes), 1001-frame rollout
bash scripts/wan2.2/predict_causal_action_dmd_mix.sh

For live play, examples/wan2.2/demo.py serves a page on port 5002 that streams the rollout chunk by chunk. Keep that tab in the foreground, since background tabs throttle playback to about 1 FPS. Actions are a JSON list where each segment ends at a frame number and sets w s a d shift ctrl, an unused slot, then mouse look in degrees per frame:

[
  [80, "1 0 0 0 0 0 0 0 0"],
  ""
]

That one walks forward for 80 frames. The last string lists frames where space is held.

Why it matters for game builders

  • Real game input: WASD, shift, ctrl, space and mouse look map straight onto a standard first-person controller
  • Streaming, not rendering: causal chunks mean the model reacts as you press, instead of rendering a finished clip
  • Engine-made data: the Capsaicin recipe shows how to capture action-labelled training clips from your own 3D scenes, and AMD released both the dataset and the training stages
  • Fully open stack: weights, inference code, training code and data are all out, which is rare among world models

The limits are real, though. The resolution is low at 352×640, the real-time figure is measured on a data-center MI350X, and the repo targets ROCm first. AMD also says 2.0 is still under active development.

Original Source

https://rocm.blogs.amd.com/artificial-intelligence/Micro-World-V2/README.html

https://huggingface.co/amd/Micro-World-2

https://github.com/AMD-AGI/Micro-World

https://huggingface.co/datasets/amd/Micro-World-Capsaicin

Conclusion

Micro-World 2 turns AMD's earlier offline world model into something that streams frames while you hold down W. It's low-resolution and tuned for AMD data-center hardware, but it ships the whole kit, from engine-captured training data to a 4-step student, so studios can study or retrain it on their own scenes.

—Titus