HunyuanVideo

HunyuanVideo is a video generation model developed by Tencent, released on 3 December 2024 with over 13 billion parameters — at launch the largest open-source video generation model. The code was published on GitHub and the weights (including FP8 quantized versions) on Hugging Face, under the Tencent Hunyuan Community License Agreement.

The model is based on a diffusion transformer architecture with a hybrid “dual-stream to single-stream” design: in the dual-stream phase, video and text tokens are processed independently, and in the single-stream phase they are concatenated for multimodal fusion. Text encoding is handled by a decoder-only multimodal large language model (MLLM) with a bidirectional token refiner, and compression is performed by a causal 3D VAE with compression ratios of 4 (temporal), 8 (spatial) and 16 (channel).

HunyuanVideo generates video from text (text-to-video), while a separate HunyuanVideo-I2V model adds image-to-video animation. It supports resolutions from 540p to 720p across various aspect ratios (9:16, 16:9, 4:3, 3:4, 1:1) and clips of 129 frames (e.g. 720×1280, 129 frames).

According to its technical report, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6 and three top-performing Chinese video generation models. It was built through careful data curation, advanced architectural design and progressive training scaling; Tencent’s goal was to bridge the gap between the closed-source and open-source communities.