Alaya Lab has released Evoke, an open-source interactive world model that generates playable, real-time video worlds from a single image. The project, presented as a three-step, CFG-free interactive world model, is described by the team as state of the art on the WBench benchmark.
How it works. The input is an image plus joystick movements, and the model outputs video that responds to those movements almost instantly. Users can steer a variety of vehicles, from a snowmobile to other drivable machines, and create events such as a volcano eruption. Text prompts can also add elements to a scene, like balloons or aurora lights.
Availability and licensing. Evoke is fully open source, and the code button on the project page contains instructions to download and run it locally. The team also released the models for every stage of training. The final model is around 57GB in size, which means a decent high-end GPU is required. The project is released under the Apache 2.0 license, which carries minimal restrictions.
Evoke specs:
Type: interactive world model (image + joystick + text prompt to video)
License: Apache 2.0, fully open source
Model size: final model around 57GB
Hardware: high-end GPU recommended
Status: open-source release, no consumer product announced
Community
0 reader comments
No comments yet. Join the conversation with a useful, respectful response.