What is fasteditwam?

FastEditWAM is a research project for training robots to perform tasks by predicting what a scene will look like next. It is a modified, smaller version of an existing project called FastWAM, adapted to work with an image editing model named Z-Image. The people behind it noticed that predicting many future video frames during training takes a lot of time and is not actually necessary. Predicting just one frame still gives the model enough visual information to learn from, while making the training process faster. The project swaps in Z-Image as the component that handles images, adds the configuration files needed to run it in a few different sizes, and precomputes text descriptions of tasks using a language model called Qwen3 so they do not need to be recalculated every time during training. It reuses the same LIBERO benchmark testing setup that the original FastWAM project used. LIBERO is a simulated environment for testing how well a robot can complete household style tasks based on instructions. So far the project has only been tested in this LIBERO simulation, not on a physical robot, though the author believes the same approach should also work on real robots. Results shown in the README compare FastEditWAM against the original FastWAM across several LIBERO task categories, and the two perform close to each other, with FastEditWAM scoring slightly better on one category and slightly lower on the others. Getting the project running involves installing the package, downloading a specific Z-Image model checkpoint, preparing the LIBERO dataset in a particular format, and separately installing the official LIBERO simulator, which is not included in this repository. From there the README walks through precomputing text embeddings, training on a multi GPU machine, and running evaluation scripts against the trained checkpoints. This is a project for people already working in robot learning research rather than a general purpose tool, and it comes with a citation entry for anyone using it in their own work.