EgoX is a novel framework for generating egocentric (first-person) videos from a single exocentric (third-person) video input. It addresses realistic viewpoint transformation while maintaining temporal consistency and scene structure. Designed for researchers in egocentric video synthesis, EgoX offers a powerful tool for creating immersive first-person perspectives by leveraging external observations and egocentric priors.

How It Works

The framework builds upon large-scale video diffusion models trained on the Ego-Exo4D dataset. EgoX employs a unified conditioning strategy integrating spatial and channel information within latent representations for realistic viewpoint transformation. A key advantage is its lightweight adaptation mechanism using LoRA-based fine-tuning, significantly reducing customization computational burden.

Quick Start & Requirements

• Installation: Requires Python 3.10, CUDA 12.1+, and compatible PyTorch. Installation involves conda environment setup, PyTorch installation, and pip installing dependencies.
• Hardware: Substantial GPU VRAM is mandatory: ≥ 80GB for inference, ≥ 140GB for training.
• Model Weights: Pretrained Wan2.1-I2V-14B and EgoX LoRA weights must be downloaded from Hugging Face/Google Drive.
• Inference: Quick testing uses example data via shell scripts (scripts/infer_itw.sh, scripts/infer_ego4d.sh). Custom data inference requires specific directory structures and metadata preparation.
• Links: Teaser Video: `https://github.com/user