Pyramid Flow: Video Generator — What It Does and What You Need to Run It

Pyramid Flow is an open source video generator that solves the text-to-video task. The model uses a pyramidal approach to flow matching and, as described by the developers, makes it possible to generate video on GPUs with a small amount of VRAM.

What it does

• Generate video from a text description (text-to-video).
• Run in Multi-GPU mode — use several GPUs for acceleration.
• Offload computations to the CPU (CPU offloading) to fit within available VRAM.

The authors describe the model as "Pyramidal Flow Matching for Efficient Video Generative Modeling" and note that it was presented at ICLR 2025.

What you need to run it

• Platform: the developers mention the MPS backend; the code is written in Python, the recommended version is 3.8.10.
• Memory: as described by the author, running it requires less than 8 GB of GPU VRAM. When using multiple GPUs and CPU offloading, this value can be even lower.
• Framework: the model is integrated with the diffusers library, and the weights are available on Hugging Face. The largest weights file is 7.8 GB, and the total size of all files is 27.2 GB.
• License: the code is distributed under the MIT license, and the model weights are under the Stability AI Community License.
• Local launch: possible, the source code is open in the repository github.com/jy0205/Pyramid-Flow. The last code change is dated December 21, 2024.

Who it suits

The tool will be useful to researchers and developers who want to understand how modern video generative models work or adapt them to their own tasks — especially since the code is open and written in Python using diffusers. The ability to run on GPUs with 8 GB of VRAM and Multi-GPU support make the model accessible to those who do not have server accelerators with a large amount of memory.

It is not suitable for those looking for a ready-made cloud service — "upload text, get video" — without diving into environment setup and working with the command line. It is also worth noting that the model weights have a separate license, different from MIT — this may matter for commercial use.

Pyramid Flow is one example of how research models are gradually lowering the barrier to entry for video generation. You can learn more about the project on the website pyramid-flow.github.io, and download the model weights from huggingface.co/rain1011/pyramid-flow-sd3.