Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Apple Open-Sources FastFrames-S and FastFrames-L Video Models

AI summary

Apple has open-sourced the 1.3B-parameter FastFrames-S and 14B-parameter FastFrames-L models for low-latency 480p video generation. Developers can download the model weights, code, training tools, and evaluation resources; the weights are licensed for noncommercial research use. Apple reported that FastFrames-S achieved 36-millisecond generation latency on one Nvidia B200 GPU and ran about 7.6 times faster than its Causal Forcing++ base model. Apple also reported applying the acceleration method to LingBot 1.3B and RealWonder, with reported response-speed gains of about 16 and 22 times, respectively.

In-depth insight

The release gives video researchers and interactive-application developers an open implementation with model weights, code, training tools, and evaluation resources. The reported acceleration method was also applied to external models, providing a potentially reusable approach for speeding up interactive video generation.

智东西 编译 | 孙小雯 编辑 | 程茜

智东西11月18日消息,今天,苹果开源面向实时交互的视频生成模型FastFrames系列,包含13亿参数的FastFrames-S和140亿参数的FastFrames-L。两款模型面向视频生成研究人员和交互应用开发者,主打低延迟480P视频生成。

据苹果公布的测试,FastFrames-S在单张英伟达B200 GPU上的生成延迟为36毫秒,在保持与其基础模型Causal Forcing++相近生成质量的同时,速度达到后者的约7.6倍。FastFrames-L与同样基于通义万相Wan2.1-14B开发的实时视频模型Krea-Realtime生成效果相当,单卡速度约为后者的10.6倍;使用8张B200后,延迟进一步降至45毫秒。

苹果开源两大视频生成模型!单卡延迟36毫秒,技术秘籍还能给别家模型提速

#Apple#FastFrames-S#FastFrames-L#Causal Forcing++#Wan2.1-14B#Krea-Realtime#LingBot 1.3B#RealWonder#Nvidia B200#Hugging Face

Original · 10 min readRead the original
Group chat

No comments yet. Start the conversation.