WorldWander is the official PyTorch codebase for the paper "WorldWander: Bridging Egocentric and Exocentric Worlds in Video Generation," which has been accepted to ECCV 2026. The project focuses on the problem of conversion between different viewpoints in video generation, especially seamless video translation between first-person (egocentric) and third-person (exocentric) viewpoints.
In terms of purpose, WorldWander aims to bridge the viewpoint gap between egocentric and exocentric worlds. This capability has potential value for directions such as filmmaking, embodied AI, and world models.
In terms of main capabilities, WorldWander is an in-context learning framework built on an advanced video diffusion Transformer, integrating two key designs: (i) In-Context Perspective Alignment and (ii) Collaborative Position Encoding, for efficient modeling of cross-view synchronization. In addition, the project also releases the EgoExo-8K dataset, which contains synchronized egocentric-exocentric triplets from synthetic scenes and real scenes.
In terms of usage, the project provides two paths: inference and training. Inference part: WorldWander is trained based on the Wan2.2-TI2V-5B model, using 4 H200 GPUs, with a batch size of 4 per GPU; the project provides LoRA checkpoints for different tasks on Hugging Face. Users can download the corresponding checkpoint and specify the corresponding configuration file for inference. The official example script is provided, and the parameter ckpt_path needs to be updated to the path of the downloaded checkpoint. It is recommended to run on a GPU with 80GB of VRAM to avoid out-of-memory. Training part: users can train on custom datasets. They need to first adjust parameters such as first_video_root, third_video_root, and ref_image_root in the corresponding configuration file, and if necessary, modify the CustomTrainDataset class in dataset/custom_dataset.py according to the properties of their own dataset. The official training script is also provided.
In terms of applicable scenarios, according to the project description, this framework is suitable for scenarios requiring cross-view video generation and conversion, such as filmmaking, embodied AI, and world models.
In terms of limitations, the materials do not provide specific performance metrics, the maximum supported resolution or duration, license information, or other hardware requirements besides the recommendation of 80GB VRAM. The project code partially draws on DiffSynth-Studio and Wan2.2.
