asdfo123/forgewm

ForgeWM is an open-source Python framework for training an AI model that can generate a playable, interactive Minecraft-like game world in real time. The model watches a starting image of a game scene, accepts keyboard and mouse inputs, and generates video frames that show what the world should look like as you move through it. The goal is to produce convincing, continuous game video at a speed fast enough to feel interactive, responding correctly to your actions. This kind of AI is called a world model. Rather than rendering a game world with traditional 3D graphics, it generates the visuals directly using a trained neural network. ForgeWM brings together three existing research components: a video generation backbone from a project called Matrix-Game 2, open Minecraft gameplay recordings from a dataset called GameFactory, and a training method called Causal Forcing that compresses the model down to only four generation steps so it runs fast enough for real-time use. The README explains why this project exists: Matrix-Game 2 released its trained model but not the training code or data, and other open alternatives focused on camera movement in general videos rather than discrete keyboard and mouse controls in games. ForgeWM fills that gap by providing the complete training pipeline, the data, and the model checkpoints, all publicly on Hugging Face. Training runs across four stages on eight high-end GPUs and takes a total of roughly 22,000 training steps across the stages. Pre-encoded training data is available for download at about 89 GB. Running inference on a trained model requires only a single GPU and a starting image, and you specify which action to simulate. The README includes a comparison table showing ForgeWM produces quality close to Matrix-Game 2 while fixing a specific visual artifact present in the original model. The project is licensed under Apache 2.0.