Pyramid Flow

Pyramid Flow is an open-source text-to-video generation framework that produces high-quality, temporally consistent videos from natural language prompts. Built on a pyramid flow matching architecture, it models video as a hierarchical sequence of latent representations, enabling efficient generation of long and coherent clips. The system supports multi-stage generation, starting from low-resolution, coarse motion and progressively refining spatial details and temporal dynamics. Pyramid Flow integrates flow-based generative modeling with diffusion-style refinement, allowing controllable and stable video synthesis.

Key capabilities include generating diverse scenes, actions, and camera motions directly from text, handling both simple and complex prompts. The framework is designed to scale to higher resolutions and longer durations while maintaining consistency across frames. It supports sampling strategies that balance quality and speed, making it suitable for experimentation and research. Use cases include content prototyping, visual storytelling, simulation of dynamic scenes, and academic research on generative video models. The project provides pretrained models, inference code, and reproducible training pipelines, enabling researchers and developers to benchmark, extend, or adapt the approach to custom datasets and domains. Pyramid Flow emphasizes transparency and reproducibility, with detailed documentation, open-source implementation, and evaluation on standard text-to-video benchmarks.