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

UniAVC is an open-source video generator from the HKUST LONG Group research team. It handles text-to-video and, judging by the repository tags, also touches on related scenarios: audio-to-video, video-to-audio, joint audio and video generation and editing, and video editing within a single model.

What it does

The authors describe the project as a unified model for creating and editing audio and video. The repository lists the tags: audio-to-video, joint-audio-video-editing, joint-audio-video-generation, unified-model, video-editing, video-generation and video-to-audio. This suggests the model is designed not only for generating video from text but also for working with the audio track — both creating it and editing it jointly with the video.

The fact sheet has no detailed breakdown of the architecture or output examples, so the quality of the result can only be judged by the stated task and the project's focus.

What you need to run it

The code is written in Python. The authors tested the environment on Python 3.10, CUDA 12.8 and PyTorch 2.8. The platform, as described by the author, is CUDA, meaning an NVIDIA graphics processor is required to run it.

The largest model weights file is 43 GB, and this same value is listed as the total size of all model weights files. As described by the author, the model requires substantially more VRAM than a single 80 GB GPU provides — verbatim: "speed and needs substantially more VRAM than a single 80 GB GPU provides at the". This is the developer's claim, not an independent measurement.

The code is distributed under the CC BY-NC 4.0 license. Its terms permit reproducing, distributing and adapting the code for non-commercial purposes only. When distributing, you must retain attribution, the license notice and indicate any changes made. Imposing additional restrictions on recipients is prohibited.

The model weights are distributed under the LTX-2 Community License. The model may be freely used, modified and distributed for any purpose if the company's annual revenue does not reach a certain threshold. Above it, a paid commercial license is required. Military uses, training competing models, creating competing products and a number of other restrictions from Attachment A are prohibited. Derivative works must be distributed under this same license.

Who it suits

The project will interest researchers and developers working with generative video and audio models who have access to powerful CUDA hardware. The open source allows studying the implementation and adapting it for non-commercial tasks.

For commercial projects, the model is suitable with caveats: the code license prohibits commercial use, and the weights license imposes restrictions on companies with revenue above the established threshold. For an ordinary user without a specialized graphics processor with a large amount of VRAM, running it will be difficult.

UniAVC is an open-source research tool focused on unified work with video and audio. Its practical application runs up against high VRAM requirements and separate licenses for the code and the weights, which should be taken into account when planning a project.