Latte: video generator — what it does and what you need to run it

Latte is an open source video generator built on the Latent Diffusion Transformer architecture. The developers position it as a research tool for synthesizing video from text descriptions, published as part of the TMLR 2025 conference.

What it does

• Generates video based on a text description (text-to-video).
• Uses the Latent Diffusion Transformer approach — working with a latent representation of frames instead of pixel-by-pixel generation.
• The authors provide pretrained models available for download and local running.
• The code is open source under the Apache-2.0 license, permitting use and modification in commercial and non-commercial projects.

What you need to run it

• Platform: as described by the authors, CUDA is required. However, the documentation notes that to run pretrained models locally on a CPU, you can remove the cudatoolkit and pytorch-cuda dependencies from the requirements file.
• Language: Python.
• License: Apache-2.0.
• Repository: github.com/Vchitect/Latte.
• Currency: the last code change was recorded on October 30, 2025.

Who it is for

The tool is aimed primarily at researchers and developers familiar with Python and the PyTorch ecosystem. The open source code and Apache-2.0 license make it a convenient base for experiments with video diffusion, fine-tuning models, or embedding into your own pipelines. The ability to run on a CPU, albeit with caveats, lowers the entry barrier for those without a powerful GPU.

For those looking for a ready-made application with a graphical interface or a cloud service without environment setup, Latte is not suitable — it is essentially a codebase for local deployment.

The authors do not specify concrete requirements for VRAM or system RAM, so before running, it is worth accounting for the typical appetites of diffusion models and testing on your own hardware.