Tencent open-sources the HunyuanVideo generation model

Background

Video generation models create moving images from text descriptions or other inputs. HunyuanVideo is Tencent's entry in this area, and Tencent presented it as the largest open-source model of its kind at the time of release ^[1].

Text-to-video models are usually built on diffusion techniques: they start from noise and refine it over many steps, and they must keep motion consistent from frame to frame, which makes them much more demanding than image generation. Open-source releases let researchers and developers run these models on their own hardware instead of relying only on a commercial service.

What happened

On 3 December 2024 Tencent released HunyuanVideo's inference code and model weights ^[1]. The accompanying technical report was published on arXiv the same day ^[2].

Key details

With over 13 billion parameters, Tencent said it was the largest open-source video generation model ^[1]. In Tencent's professional human evaluation, the model ranked first overall against several closed commercial video models, including Gen-3 alpha and Luma 1.6 ^[1]. Running it locally needs at least 60GB of GPU memory, according to Decrypt ^[3].

Parallel inference code followed on 7 December, then Diffusers support on 17 December and FP8 weights on 18 December 2024 ^[1].

Motion quality, which the Tencent evaluation singled out, is one of the harder aspects of video generation, because objects must move smoothly and plausibly across frames. Evaluation results of this kind also depend on the prompts, the evaluators and the rival models chosen.

Impact

The release made the code and weights of a large open video model available to outside users. The memory requirement, however, means that the model is practical mainly on high-end GPUs ^[1][3].

Hardware requirements of this kind are one reason large open models are used mainly by laboratories and companies with dedicated computing resources.