MIRA is a real-time world model for Rocket League: a 5B-parameter latent diffusion model generates video frame by frame based on the actions of all four players, so a complete 2v2 match can be played within the model at 20 FPS on a single GPU. It can be played in real time at mira-wm.com.
This is the official code release for the technical report "MIRA: Multiplayer Interactive World Models with Representation Autoencoders," completed through a collaboration between General Intuition, Kyutai, and Epic Games.
Purpose and capabilities: MIRA is aimed at multiplayer interactive world modeling, capable of generating game visuals in real time based on the synchronized actions of four players, and supports complete 2v2 matches within the model. The technical approach is a latent diffusion model combined with representation autoencoders.
Data: Each sample is a 4-second window from a 2v2 match, collected from the synchronized perspectives of all four players. For each frame of each perspective, the dataset provides the video frame, the player's keyboard actions, and the game state (ball, vehicles, and score). The complete dataset is hosted on the HuggingFace Hub (see the dataset card). Interactive exploration is available (4-player grid, keyboard overlay, top-down global view using game state).
Usage: Requires pixi and torch >= 2.8 (installed automatically). The entry points under scripts/ are Hydra applications—any configuration key can be overridden via key=value. Multi-GPU runs are done via torchrun; single-GPU runs are also possible. Codec training uses a frozen DINOv3-L/16 encoder, whose weights are gated by Meta. Download dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth from the DINOv3 page and set RS_DINO_WEIGHTS_DIR=/path/to/weights. It is not needed for world model training and inference.
License: Apache License 2.0—see LICENSE.
Applicable scenarios: game AI research, world model research, interactive simulation, real-time video generation, multi-agent action-conditioned generation.
Limitations: Codec training relies on Meta-gated DINOv3 weights, which must be downloaded and configured manually; world model training and inference do not require these weights. Real-timeplay requires access to mira-wm.com or setting up the environment yourself. Apart from "20 FPS on a single GPU," specific hardware requirements have not yet been provided.
