Latte

Why we included this project

Latte is the official PyTorch implementation of a latent diffusion transformer for generating video from text prompts. The repository ships pre-trained checkpoints trained on FaceForensics, SkyTimelapse, UCF101, and Taichi-HD, plus a higher-resolution text-to-video model, so you can judge real output quality before committing to training anything from scratch. Beyond sampling scripts, it includes distributed training and evaluation code, which makes it a practical starting point for teams adapting transformer-based diffusion to their own video tasks. The model also plugs into the diffusers library, where 4/8-bit quantization noticeably cuts GPU memory during inference. Documentation covers architecture variants and ablation choices, useful when you want to understand why one design decision beats another.