videogan

Why we included this project

This is the reference implementation behind the NIPS 2016 paper that showed adversarial networks could hallucinate short, plausible video clips from scene categories like beaches, golf courses, and train stations. The Torch7 codebase builds on DCGAN and is a compact, readable starting point for researchers and students working on generative video models: it separates training from conditional generation from a static image, and the generate script exposes intermediate layers such as the motion mask and background. The authors include pre-trained models and a stabilization pipeline that flattens video into vertically concatenated JPEG frames for efficient loading. It is not a production-ready tool, but it is a useful historical baseline for understanding how early video GANs modeled scene dynamics and for reproducing the paper's results.