Video super-resolution without the flicker

What it does TecoGAN is a temporally coherent GAN for video super-resolution. It takes low-resolution video and generates high-resolution frames while keeping fine details—such as armor mesh, lizard scales, or spider markings—stable over long sequences. The repository includes inference code, training pipelines, and pre-trained models from the published SIGGRAPH work.

The interesting bit The authors use a spatio-temporal discriminator to guide the generator toward frame-to-frame consistency, so hallucinated textures persist instead of flickering. They also provide custom temporal metrics, tOF and tLP, alongside PSNR and LPIPS to measure that coherence rigorously.

Key highlights

• Research release shipping both training and inference code, plus pre-trained models
• Generates fine textures that remain coherent across extended video sequences
• Includes temporal evaluation metrics (tOF, tLP) implemented in PyTorch
• Training data downloader pulls from online video sources, with fallback replacements for offline content
• TensorBoard integration renders GIF summaries of training progress

Caveats

• Evaluation currently requires an Nvidia GPU with CUDA; the README notes this is needed for test case 2
• 36 of the 308 training sequences are no longer online, so the data prep script substitutes replacements
• The README specifies TensorFlow 1.8+ and youtube-dl for data preparation

Verdict Worth exploring if you need a published baseline for temporally stable video super-resolution, or want to benchmark against the ACM TOG paper. Skip it if you need a modern, production-ready stack—the dependencies and hardware requirements reflect its research origins.