阿里万相视频生成大模型宣布开源:8.2GB 显存就能跑,测试超越 Sora

阿里万相视频生成大模型宣布开源:8.2GB 显存就能跑,测试超越 Sora

作者: 发表时间:2025-08-14 10:18:54
==> 1 ok ==> 2 ok ==> 3 ok ==> 4 ok ==> 5 ok ==> 6 ok ==> 7 ok ==> 8 ok ==> 9 ok ==> 10 ok ==> 11 ok ==> 12 ok ==> 13 ok ==> 14 ok ==> 15 ok ==> 16 ok ==> 17 ok ==> 18 ok ==> 19 ok ==> 20 ok ==> 21 ok ==> 22 ok ==> 23 ok ==> 24 ok ==> 25 ok ==> 26 ok ==> 27 ok ==> 28 ok ==> 29 ok ==> 30 ok ==> 31 ok ==> 32 ok ==> 33 ok ==> 34 ok ==> 35 ok ==> 36 ok ==> 37 ok ==> 38 ok ==> 39 ok ==> 40 ok ==> 41 ok ==> 42 ok ==> 43 ok ==> 44 ok ==> 45 ok ==> 46 ok ==> 47 ok ==> 48 ok ==> 49 ok ==> 50 ok ==> 1 ok ==> 2 ok ==> 3 ok ==> 4 ok ==> 5 ok ==> 6 ok ==> 7 ok ==> 8 ok ==> 9 ok ==> 10 ok ==> 11 ok ==> 12 ok ==> 13 ok ==> 14 ok ==> 15 ok ==> 16 ok ==> 17 ok ==> 18 ok ==> 19 ok ==> 20 ok ==> 21 ok ==> 22 ok ==> 23 ok ==> 24 ok ==> 25 ok ==> 26 ok ==> 27 ok ==> 28 ok ==> 29 ok ==> 30 ok ==> 31 ok ==> 32 ok ==> 33 ok ==> 34 ok ==> 35 ok ==> 36 ok ==> 37 ok ==> 38 ok ==> 39 ok ==> 40 ok ==> 41 ok ==> 42 ok ==> 43 ok ==> 44 ok ==> 45 ok ==> 46 ok ==> 47 ok ==> 48 ok ==> 49 ok ==> 50 ok

本站 2 月 25 日消息,阿里云今日宣布旗下视觉生成基座模型万相 2.1(Wan)开源。

此次开源采用最宽松的 Apache2.0 协议,14B 和 1.3B 两个参数规格的全部推理代码和权重全部开源,同时支持文生视频和图生视频任务,全球开发者可在 Github、HuggingFace 和魔搭社区下载体验。

据介绍,14B 万相模型在指令遵循、复杂运动生成、物理建模、文字视频生成等方面表现突出,在评测集 VBench 中,万相 2.1 以总分 86.22% 的成绩超越 Sora、Luma、Pika 等国内外模型,位列榜首。1.3B 版本测试结果不仅超过了更大尺寸的开源模型,甚至还接近部分闭源模型,同时能在消费级显卡运行,号称“仅需 8.2GB 显存就可以生成 480P 视频”,适用于二次模型开发和学术研究。

万相视频模型架构图

在算法设计上,万相基于主流 DiT 架构和线性噪声轨迹 Flow Matching 范式,研发了高效的因果 3D VAE、可扩展的预训练策略等。以 3D VAE 为例,为了高效支持任意长度视频的编码和解码,万相在 3D VAE 的因果卷积模块中实现了特征缓存机制,从而代替直接对长视频端到端的编解码过程,实现了无限长 1080P 视频的高效编解码。此外,通过将空间降采样压缩提前,在不损失性能的情况下进一步减少了 29% 的推理时内存占用。

万相团队的实验结果显示,在运动质量、视觉质量、风格和多目标等 14 个主要维度和 26 个子维度测试中,万相均达到了业界领先表现,并且斩获 5 项第一。

本站附开源地址:

    Github:https://github.com/Wan-Video

    HuggingFace:https://huggingface.co/Wan-AI

    魔搭社区:https://modelscope.cn/organization/Wan-AI

相关文章