What is giga-world-1?

GigaWorld-1 is a research project that builds world models for evaluating robot control policies. A world model is a system that learns to predict what a scene will look like as a robot acts in it, so researchers can test and compare robot behavior without needing a physical robot or a real environment every time. The project releases open code for training these models, running inference on them, preparing training data, converting checkpoints, and merging LoRA adapters. It ships two model sizes, a smaller Nano version and a larger Pro version, both trained in two stages: a first stage that teaches the model to follow controllable inputs, and a second stage that distills the model down so it runs faster with fewer computation steps. Training code supports DeepSpeed for splitting large models across multiple GPUs. Once trained, the models can generate video predictions from either an image or a text description, producing short rollouts of about 33 seconds at 10 frames per second. The data preparation pipeline converts robot demonstration data from the LeRobot format into GigaWorld's own format, adding automatic captions and depth information using other open models. A toy dataset is provided on Hugging Face so people can try the pipeline without needing large-scale robot data. The repository also introduces WMBench, a benchmark with detailed metrics and a leaderboard for scoring how well a world model predicts realistic outcomes, judged partly by other AI models. Some pieces are still coming: distilled versions of the released weights, reinforcement-learning-based training for better understanding of 3D scenes, model weights for other domains beyond robotics, and a faster training and inference stack are all marked as upcoming in the project roadmap. This is aimed at machine learning researchers and engineers working on robotics or generative video models, not general software developers. It requires a Linux machine with capable GPUs, Python 3.10 or newer, and PyTorch. The code is released under the Apache 2.0 license.