腾讯开源混元视频模型

2024 年 12 月 3 日，腾讯混元把 HunyuanVideo 的代码与权重放上 GitHub 和 Hugging Face。任何人都可以看到采样命令：输入一句提示，设置 720×1280、129 帧与推理步数，再等待结果写入目录。与只能排队调用的产品相比，这是一次实质性的开放；但从 git clone 到第一段视频之间，还隔着一台大多数个人用户并不拥有的机器。

项目说明称模型规模超过 130 亿参数。生成 720p、129 帧视频时，单卡峰值显存约需 60GB，官方测试使用 80GB GPU；较低的 544×960 设置仍约需 45GB。FP8 权重可以再节省约 10GB，多卡并行和 CPU offload 也能调整门槛。开放权重让运行成为可能，却没有让视频生成突然像本地文生图一样轻便。

HunyuanVideo 的管线解释了资源花在哪里。3D VAE 先把图像和视频压进时空潜变量，文本提示由多模态大语言模型编码，扩散 Transformer 再从噪声中生成视频潜变量。架构先分别处理文本与视频 token，随后将它们合流，以增强语义与画面的互动。129 帧不是 129 张彼此独立的图片；模型必须同时维护运动、外观与时间连续性。

仓库还提供提示改写模型，把用户短句转成更适合视频生成的描述。Normal 模式偏向保留意图，Master 模式会补充构图、光线与镜头运动以追求画面质量，但官方也承认这种扩写有时会丢失语义细节。这是一项常被漂亮样片隐藏的取舍：模型看似更理解提示，也可能只是另一个模型先替用户改写了提示。

腾讯报告使用 1,533 条提示和专业评审进行比较，并称 HunyuanVideo 在若干闭源模型对照中表现领先。报告同时注明，评测使用的是高质量版本，与当时公开的快速版本不同。这个限定词非常关键：发布方评测不能直接代表下载者在公开权重、自己的硬件和默认参数下会得到同样结果。开放的价值首先是允许外部真正检查这种差异。

后来的图生视频、Avatar 与 1.5 版本扩展了混元视频家族，但首发已经留下一个清楚的变化。大型视频模型不再只能以 API 名称存在，它可以被下载、拆解、量化和接入 Diffusers 等社区工具；与此同时，60GB 显存把“可获得”与“容易使用”分得很开。HunyuanVideo 打开的不是一扇无门槛的门，而是一间终于允许外部进入、仍堆满重型设备的实验室。

Tencent Hunyuan placed HunyuanVideo’s code and weights on GitHub and Hugging Face on 3 December 2024. Anyone could read the sampling command: provide a prompt, set a 720-by-1280 frame size, request 129 frames and a number of inference steps, then wait for a file to appear. Compared with a product available only through a queue, this was substantive openness. Between git clone and the first generated clip, however, stood a machine most individuals did not own.

The project described a model with more than thirteen billion parameters. Generating a 720p, 129-frame clip required about 60GB of peak GPU memory for a batch of one, and Tencent tested on an 80GB card. Even the lower 544-by-960 setting required about 45GB. FP8 weights could save roughly another 10GB, while multi-GPU parallelism and CPU offload offered other tradeoffs. Open weights made execution possible; they did not make large video generation suddenly as light as an ordinary local image model.

The pipeline explained where the resources went. A 3D VAE compressed images and video into a spatiotemporal latent space. A multimodal large language model encoded the text prompt. A diffusion transformer generated video latents from noise before the VAE decoded them. The architecture first processed text and video tokens in separate streams, then joined them to increase interaction between semantics and imagery. One hundred twenty-nine frames were not 129 independent pictures. Motion, appearance, and temporal continuity had to be maintained together.

The repository also supplied a prompt-rewrite model that expanded a user’s short instruction into language better suited to video generation. Normal mode aimed to preserve intent; Master mode added composition, lighting, and camera movement to pursue visual quality. Tencent noted that the latter could occasionally lose semantic detail. This was a tradeoff often hidden by attractive samples: the video model might appear to understand a prompt better because another model had rewritten the prompt first.

Tencent reported an evaluation using 1,533 prompts and professional raters, with HunyuanVideo outperforming several closed systems in its comparisons. The report also stated that the evaluation used a high-quality version different from the fast version initially released. That qualifier was essential. A publisher comparison could not guarantee that a downloader using public weights, different hardware, and default settings would reproduce the same result. The first value of openness was the ability to investigate that gap directly.

Later image-to-video, Avatar, and 1.5 releases expanded the family, but the first release had already made a clear change. A large video model no longer existed only as an API name. It could be downloaded, inspected, quantized, and integrated into community tooling such as Diffusers. At the same time, the 60GB requirement separated obtainable from easy to use. HunyuanVideo did not open a frictionless door. It opened a laboratory that outsiders could finally enter, while leaving the heavy equipment inside.