{"id":29620,"date":"2025-02-26T11:27:23","date_gmt":"2025-02-26T03:27:23","guid":{"rendered":"https:\/\/www.1ai.net\/?p=29620"},"modified":"2025-02-26T11:27:23","modified_gmt":"2025-02-26T03:27:23","slug":"%e9%98%bf%e9%87%8c%e4%b8%87%e7%9b%b8%e8%a7%86%e9%a2%91%e7%94%9f%e6%88%90%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%ae%a3%e5%b8%83%e5%bc%80%e6%ba%90%ef%bc%9a8-2gb-%e6%98%be%e5%ad%98%e5%b0%b1%e8%83%bd%e8%b7%91","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/29620.html","title":{"rendered":"Ali Wanphase Video Generation Big Model Announced Open Source: 8.2GB Video Memory Can Run, Tests Surpass Sora"},"content":{"rendered":"

Alibaba Cloud<\/a>announced on February 25th that its visual lifeinto a base modelgazillion-phase<\/a> 2.1 (Wan) Open Source.<\/strong><\/p>\n

\"5e25399fj00ss9tj1009rd000u000kup\"<\/p>\n

This open source uses the most lenient\u00a0Apache 2.0 protocol<\/strong>The newest version of the program is the \"14B\" and \"1.3B\" parameter specifications, which are open source and support both text-generated video and graph-generated video tasks, and can be downloaded from Github, HuggingFace, and the Magic Hitch community.<\/p>\n

It is reported that the 14B Wanphase model excels in command following, complex motion generation, physical modeling, and text and video generation in the review set VBench.With a total score of 86.22%, Manphase 2.1 outperforms domestic and international models such as Sora, Luma, and Pika.<\/strong>Version 1.3B not only outperforms larger open-source models and even approaches some closed-source models, but also runs on consumer graphics cards, claiming \"Generate 480P video with only 8.2GB of video memory<\/strong>\" for secondary model development and academic research.<\/p>\n

\"b026c831j00ss9tjl0092d000u000ghp\"<\/p>\n

In terms of algorithm design, Maxthon has developed efficient causal 3D VAE and scalable pre-training strategies based on the mainstream DiT architecture and the Flow Matching paradigm for linear noise trajectories. Taking 3D VAE as an example, in order to efficiently support the encoding and decoding of videos of arbitrary length, Maxthon implements a feature caching mechanism in the causal convolution module of 3D VAE, which replaces the direct end-to-end coding and decoding process of long videos, and realizes the efficient coding and decoding of unlimited-length 1080P videos. In addition, by advancing the spatial downsampling compression, the memory footprint of 29% during inference is further reduced without performance loss.<\/p>\n

The experimental results of the Wanxiang team show that in the 14 main dimensions and 26 sub-dimension tests of motion quality, visual quality, style and multi-targetingThe 10,000 phases have achieved industry-leading performance and five firsts.<\/strong>.<\/p>\n

With open source address:<\/p>\n