🤖 AI 资讯

每日 05:00 更新 · 09-17 · 主站 liuch.name ↗
全部标签 →
筛选标签:数字人 · 返回个性化推荐 · 清空筛选
AI 资讯

GenStream: Semantic Streaming Framework for Generative Reconstruction of Human-centric Media

arXiv cs.AIarXiv:2609.18634v1 Announce Type: cross Abstract: Video streaming dominates global internet traffic, yet conventional pipelines remain inefficient for structured, human-centric content such as sports, performance, or interactive media. Standard codecs re-encode entire frames, foreground and background alike, treating all pixels uniformly and ignoring the semantic structure of the scene. This leads to significant bandwidth waste, particularly in scenarios where backgrounds are static and motion is constrained to a few salient actors. We introduce GenStream, a semantic streaming framework that replaces dense video frames with compact, structured metadata. Instead of transmitting pixels, GenStream encodes each scene as a combination of skeletal keypoints, camera viewpoint parameters, and a static 3D background model. These elements are transmitted to the client, where a generative model reconstructs photorealistic human figures and composites them into the 3D scene from the original viewpoint. This paradigm enables extreme compression, achieving over 99.9% bandwidth reduction compared to HEVC for the continuous data stream. We partially validate GenStream on Olympic figure skating footage and demonstrate potential for high perceptual fidelity under minimal data. While acknowledging the significant computational costs shifted to the client and challenges in generalization, GenStream opens new directions in volumetric avatar synthesis, canonical 3D actor fusion across views, and personalized viewing experiences, laying the groundwork for scalable, intelligent streaming in the post-codec era.
2026-09-17 04:00:00 · 具身智能,Meta,数字人,论文
AI 资讯

They Uploaded a File and It Was Not an Image

DEV.to

Somebody uploaded a profile picture.

It was not a picture.

The endpoint checked the extension.

The file was called avatar.png,
and an extension is not a fact.

It is a claim,
typed by the person uploading.

The content type header
is the same claim,
sent by the same person.

So the checks we were proud of
amounted to asking a stranger
whether the stranger could be trusted,
and writing down the answer.

Here is the damage,
roughly in the order it arrives.

The filename went into a path.

Names can contain dots and slashes,
and a name that walks up two directories
is not exotic.

It is the first thing anybody tries.

The file landed in a directory
the web server serves,
which means we hand it back
to the next person who asks,
from our own domain,
with our own session in scope.

An SVG is not an image.

It is a document,
and documents carry script.

A PDF is a small programming language
with an unusually good reputation.

And the innocent ones
get passed to an image library
that is thirty years of C
reading a length field
chosen by somebody hostile.

Then there is size.

An archive that is two kilobytes on disk
and four gigabytes when opened
still works, because it always did.

What actually helps is dull.

Never use the name you were given.
Store the file under an identifier you generated,
and keep the original name as data,
never as a path.

Put it somewhere the web server
will not serve on its own,
and hand it back through code
that decides the type,
from a domain that is not holding the session.

Do not validate an image.

Re-encode it.
Decode it, discard the original,
write a new file out of the pixels.
Nothing hiding in the metadata survives that.

Cap the size before you read it,
not after.

Every other input you accept
is text you will parse yourself.

An upload is a file
you hand to a parser you did not write,
on behalf of somebody
you have never met.

– Serguey Asael Shinder

2026-09-17 04:36:40 · Meta,数字人,强化学习,端侧AI,招聘HR

Android Auto update lets you customize your Google Maps ‘driving avatar’ via car display

9to5Google

Google is rolling out an update for Maps on Android Auto that gives users the ability to edit their “driving avatar” directly from the car’s display.

more…
2026-09-16 17:25:00 · AI应用,Google,搜索RAG,数字人,扩散模型,招聘HR

Android Auto is finally letting you ditch that boring blue arrow straight from your dashboard

Android Authority

Android Auto in Volkswagen ID.4 Google Maps Navigation

Credit: C. Scott Brown / Android Authority
TL;DR
  • Google is rolling out the ability to change your Google Maps driving avatar directly from the Android Auto dashboard.
  • The new “Driving avatar” option lets you choose a vehicle icon and change its color without reaching for your phone.
  • We’ve spotted the feature on two devices running Android Auto 17.8.163744 and Google Maps 26.38.01.

Google is finally rolling out one of the small but useful Google Maps changes we first spotted earlier this year.

2026-09-16 07:00:59 · Google,数字人,扩散模型,招聘HR

AIGCPanel v2.4.0 绿幕视频一键换背景,歌曲也能翻唱成你的声音

开源中国绿幕换背景、把一首歌的人声换成自己的音色,这两件以前要开剪辑软件、装一串插件、再导来导去才能做完的事,现在在 AIGCPanel 里都是"上传,然后点击"。 AIGCPanel 是一站式 AI 数字人桌面应用,Windows / macOS / Linux 都能跑。v2.4.0 带来两个新工具,顺手把任务调度的老毛病一起治了:一次提交多个任务时,排队等待...
· AI应用,数字人

Your Agent Aced the Task. Will It Do It Again?

HuggingFace Blog
2026-09-15T16:00:44.002Z · 大模型,算力芯片,AI应用,开源,OpenAI,Google,Agent智能体,搜索RAG,数字人,扩散模型,强化学习,微调蒸馏,模型评测,提示工程,招聘HR,网络安全,榜单评测,论文,开发者生态

[Event Sourcing] Trying out Sekiban DCB: Implementation

DEV.to

Introduction

This article is a continuation of "[Event Sourcing] Trying out Sekiban DCB: Introduction". In the previous post, we covered setting up the DCB Native environment and its processing flow.

Initially, features like creating a Student, fetching individual and list data, and enrolling/dropping classes are already implemented. The student's state is built from recorded events. The information a student holds includes Student ID, Name, Max Class Count, and a list of currently enrolled classes.

In this post, we will add a feature to update a student's name and their maximum class limit. The Student ID and enrolled classes will remain unchanged. Decreasing the max limit below the current number of enrolled classes will be prohibited.

When updating, we don't directly overwrite the current data. Instead, we record a StudentProfileUpdated event. A Projector then processes this event to build the updated student state.

Update Feature Specifications

Item Specification
StudentId Specifies the student to update. It is not changed.
Name Required, 1 to 100 characters
MaxClassCount 1 to 10. Changing it to a value lower than the current enrollment count is not allowed.
Enrollment list Keeps the contents before the update.
Non-existing student The update is rejected.

The API receives JSON containing the student ID.

POST /api/students/update
Content-Type: application/json
{
  "studentId": "9912a519-a6f6-41fe-92bf-07b99045efff",
  "name": "John",
  "maxClassCount": 3
}

We will add the implementation in the order of Event, Decider, Command, Projector, and API.

The sample code is available on GitHub.

1. Add an Event That Represents the Update

We define the fact that the student information was updated as StudentProfileUpdated event.

using Dcb.ImmutableModels.Tags;
using Sekiban.Dcb.Events;

namespace Dcb.ImmutableModels.Events.Student;

public record StudentProfileUpdated(
    Guid StudentId,
    string Name,
    int MaxClassCount<
2026-09-16 04:50:24 · 大模型,AI应用,开源,Meta,搜索RAG,数字人,招聘HR,榜单评测,开发者生态

While Hollywood Fears an AI Future, China's Film Industry is Embracing It

SlashdotThe Los Angeles Times reports: Chen Yilong has racked up plenty of film and TV credits in his two decades as an actor, but with work getting scarcer he signed a contract in August to license the image of his face to a Chinese production studio. Chen's role will be to sit in front of a video camera and make facial expressions at the prompt of a director. A neutral stare. An angry glare. A look of surprise. Using artificial intelligence, the studio will use images of Chen's face to generate an avatar, also known as a "digital human," to star in an AI-generated movie. "If you can't beat it, join it," said Chen, 38, who works in Beijing. The Chinese film and video industry is being transformed by AI-driven storytelling, fueled by rapid advances in video generation software, and at ground zero are the so-called micro dramas that play out on millions of smartphones. Typically just a minute or two in length, the videos are devoured by Chinese audiences... And the cost of producing them has been cut drastically through the use of AI software developed by Chinese technology juggernauts including Kuaishou Technology and TikTok's global owner, ByteDance. The number of Chinese-made micro dramas surged in the first three months of this year to 128,000, according to the China Netcasting Services Assn. More than 95% of them were made with AI, the association said.... Sun Wei founded Feixiang Universe, the Shenzhen-based studio that cast Chen. The fear of accidentally "stealing someone's face" was a big reason she said she decided to license real people's likenesses for an AI production set during China's Tang Dynasty more than 1,000 years ago. Thanks to AI's training data bias, AI-generated performers tend to have similar features and share a homogenous look, Sun said. Many have flawless skin and extremely symmetrical faces, requiring her and other Chinese AI filmmakers to cast a wide net to "buy" new faces and digitize actor's expressions. Lu Beike, who directed the big-budget Chinese series "Three-Body," felt theAI-generated scenes he'd tried came out sub-standard. "Lu added that if 90% of a film was generated by AI, a real human performance could appear jarring. A real actor's expressions carried more nuance and emotion. The skin textures did not match. Sometimes the only solution was to process the live-action material until the person looked a little less real — until they fit back into the 'AI world,' Lu said." But at Sun Wei's studio, they're thrilled they can produce a 90-minute AI film for $500,000 where professional productions used to cost millions of dollars. According to the article, Sun's team works with ChatGPT, Kimi or DeepSeek to flesh out entire screenplays from a short paragraph. The screenwriter takes over, but "to produce the finished product, Sun's team uses Chinese video-generation models to input the type of scenes, backgrounds and performers they want, with the AI generating 15 seconds of footage each time." Still, the article points out that "Similar tools are being tested in the U.S., and the AI trends upending China's entertainment industry may be a harbinger for what's to come in Hollywood, said Michael Berry, a professor specializing in contemporary Chinese culture at UCLA. Already, Chinese companies that sell AI-powered video-generation software are seeking inroads in Hollywood..."

Read more of this story at Slashdot.

2026-09-15 23:34:00 · 大模型,AI应用,自动驾驶,OpenAI,DeepSeek,月之暗面,搜索RAG,数字人,扩散模型,提示工程,端侧AI,招聘HR,网络安全,开发者生态

AI ‘Actor’ Tilly Norwood Told Me That ‘All Lives Matter’

WiredThe virtual character, which is promoting its upcoming movie Misaligned, tries to evade politics by repetitively commenting on the clothes you’re wearing.
2026-09-15T18:00:09.053Z · 算力芯片,具身智能,Meta,对话助手,推理思考,数字人,扩散模型,强化学习,榜单评测

小破站砸10亿搞AI,图啥?

钛媒体

(本文作者为 AIX财经,钛媒体经授权发布)

文 | AIX财经(AIXcaiijng),作者 | 陈丹,编辑 | 魏佳

B站刚学会赚钱,就准备把一整年的利润重新押出去。

2025年,这家公司第一次实现全年GAAP盈利,净利润11.9亿元。几个月后,管理层宣布,2026年将额外增加约10亿元AI相关资本开支。

这几乎相当于B站上一年全部净利润。但放进今天的AI竞赛里,10亿元又显得拿不出手。

字节跳动今年被曝将AI基础设施预算进一步上调至2000亿元以上,9月又从近30家银行获得296亿美元贷款,大量资金继续流向芯片和数据中心;腾讯二季度资本开支达到528亿元,同比增长176%,自由现金流二十年来首次转负。

巨头正在用千亿元购买下一张技术船票,B站拿出的,只是它们一个季度花掉的很小一部分。

需要指出的是,B站并不打算参加这场最烧钱的战争。CEO陈睿在二季度财报电话会上划出了一条明确边界:“AI产业链上有很多环节我们不会参与。”钱只投三个方向:视频理解、视频分发和视频创作。

对AI行业来说,这笔钱太小,不足以买来技术领先;对B站来说,它又足够大,大到可以吃掉这家刚刚盈利公司的大部分年度利润。

B站既不想加入军备竞赛,也不能站在军备竞赛之外。

因此问题来了,一家不造大模型的内容平台,为什么仍然必须花这么多钱押AI?这笔钱究竟是在买增长,还是只是在买“不掉队”的资格?

01.10亿买的是什么?

10亿元放在整个AI行业里不算多,但对B站自己来说,是一次明显的投入升级。

2023年,B站用于设备等固定资产的资本开支约1.8亿元,2024年增至约4.8亿元,2025年约5亿元。到2026年,仅AI相关资本开支就计划额外增加约10亿元,这一增量相当于2025年全年设备投入的两倍。到二季度末,这笔预算已执行七八成,主要用于购买服务器和算力。

这也意味着B站这一轮AI投入,主要是在给现有业务补AI基础设施,让它大规模跑进搜索、推荐、内容理解和创作工具里。

这是B站2023年就定下的方向。

当年7月,“百模大战”最热时,互联网公司争夺的是通用基础模型,比参数、比训练规模、比模型能力,再把模型作为独立产品对外提供。B站也发布了自研的bilibili index大模型,但目标并不是参与这场竞争,而是让模型理解B站自己的视频和用户,服务搜索、内容审核、视频总结等内部场景。

图源 / IndexTTS2展示平台

一个月后的二季度财报电话会上,陈睿把这条路线说得更明确:B站不会参加“百模大战”。他举的例子是内容审核——如果大模型能够替代大量重复的人力,对B站来说就已经有实际价值。

此后三年,B站的AI路线基本没有偏离这个方向。在好耶科技创始人吴杰茜看来,这与其说是主动克制,不如说是现实选择。“现在才开始做(视频模型),已经太慢了,B站也不具备这个基因。”

她介绍,从零训练一个视频生成模型,再一路做到可以商用,拼的不只是钱。可灵、Seedance等模型并非短期投入的产物,背后是团队、数据和工程体系的长期积累。即便现在,视频生成模型仍以几个月为周期快速迭代。对后来者来说,增加预算可以买来算力和人才,却很难在短时间内补上已经形成的技术和工程差距。

对B站来说,更划算的办法,是让上游公司承担最昂贵的模型竞争,等能力成熟、调用成本下降,再把它们接进自己最熟悉的视频业务。

最早,AI主要待在后台,做审核、推荐和智能字幕,之后才逐渐走到用户和创作者面前。2024年,B站推出数字分身工具“必剪Studio”;2025年的“AI原声翻译”,则进一步把翻译、语音合成、字幕处理和口型模拟放进同一套流程,让一条中文视频可以更低成本地进入其他语言市场。

到今年,AI开始从辅助工具进入两个核心环节:一端理解内容,一端生产内容。

B站以中长视频为主。一条几十分钟的视频里可能同时包含人物、观点、情绪和复杂的叙事结构。传统推荐算法更多依赖标题、标签、点击和用户行为,却很难真正理解一条视频在讲什么,更难理解用户为什么会喜欢它。陈睿在二季度财报电话会上提到,新一代模型已经可以更深入地理解视频内容和用户意图。

在B站的设想中,如果这件事能成立,就不只是少雇几个审核员、少做几道字幕工序,而是可能改变搜索和推荐本身——平台可以更准确地理解自己拥有的数亿条视频,再把它们匹配给合适的人。

另一端是生产。过去,一条动画或者制作复杂的视频往往需要多人协作;现在,一个创作者借助AI,也可能完成过去需要一个小团队才能完成的工作。B站已经出现个人使用AI制作动画、获得千万级播放的案例。

今年7月,曾在腾讯混元、Anuttacon任职的曾爱玲加入B站负责AI视频生成业务,并直接向陈睿汇报。这也意味着,AI视频成为创始人直接关注的业务。

因此,对B站来说,过去它可以低成本地试用AI,把模型接进几个边缘环节;现在,当AI开始进入内容理解、分发和生产这些核心链条,B站就必须为它配置真正的算力、服务器和组织资源。

02.10亿买得到能力,买不到壁垒

这套路线已经给B站带来了一些可以量化的收益。

2026年一季度,B站广告收入25.9亿元,同比增长30%;二季度增至31.3亿元,同比增长28%。二季度,广告点击和转化综合指标CTCVR同比提高19%,搜索广告收入翻倍。管理层把其中一部分改善归因于AI对内容和用户意图理解能力的提升。

AI公司本身也在贡献增量。一季度,AI行业广告预算在B站同比增长170%,二季度仍保持翻倍增长。

内容端也是如此。二季度,B站日均投稿量同比增长28%,千粉以上创作者数量增长30%;今年5月推出的一项动画创作活动,三个月累计获得超过1.8亿次播放。

从这个角度看,B站确实吃到了“不造模型”的红利:不用承担最昂贵的技术竞赛,也能分享模型进步带来的效率提升和新增需求。

但问题也恰恰出在这里。

B站可以买到更好的视频理解能力,抖音、快手和小红书同样可以买;AI可以提高B站的广告匹配效率,也会同步提高其他平台的效率。当模型越来越接近一种通用基础设施,它首先抬高的是整个行业的效率,而不是某一家公司的独占优势。

也就是说,B站花10亿元可以买到“跟上”,却很难单靠这笔钱买到“领先”。

 

AI改变的不只是平台内部的效率,还在改变B站所处的竞争环境。

首先是内容供给。过去,视频是一种生产成本很高的内容形态。真人拍摄、动画、配音和剪辑都需要时间和人力,也因此天然限制了供给速度。AI短剧、漫剧和生成式视频出现之后,原本昂贵、低频的视频内容开始变得更便宜,也更容易批量生产。

红果便是一个直接的例子。QuestMobile数据显示,截至2026年7月,红果短剧日活达到1.68亿,已经超过爱奇艺、腾讯视频、优酷和芒果TV四家的日活总和。今年2月,红果人均单日使用时长达到125分钟;相比之下,B站今年一季度创下历史新高的日均使用时长为119分钟。

红果的增长当然不能简单归因于AI。免费模式、字节的流量体系和番茄小说的IP供给,仍然是它更直接的优势。但AI正在进一步放大这套模式的竞争力:当视频生产成本下降、供给速度加快,红果可以用更低的成本持续扩大内容池,再依靠现有的流量和IP体系把新增内容推给用户。

对B站来说,威胁并不只是“市场上的视频变多了”,而是一个已经拥有巨大流量和内容供给能力的竞争对手,可以借助AI更便宜、更快地增加视频供给,继续争夺用户每天有限的观看时间。

“我相信,不会有任何一个内容平台,面对红果飙升的用户数据不感到着急。”吴杰茜说。

竞争压力同样出现在创作者一端。

AI能够帮助UP主提高产能。一个过去需要几天甚至几周完成的视频,现在部分环节可以交给脚本生成、图片生成、配音、翻译和剪辑工具。对B站,这增加了站内供给;对创作者,它也降低了多平台分发的成本。

所以,AI带给B站的是两种同时发生的变化。

它提高平台内部的推荐、广告和创作效率,也提高整个行业的内容供给效率。前一种变化让B站愿意投入,后一种变化让B站很难选择不投入。

在吴杰茜看来,B站这10亿元与其理解为一次主动进攻,不如先把它看成一项防御性工程。这10亿元能解决的是“不掉队”,而不是建立新的护城河。

03.一张必须购买的入场券

这种防御要付出多大的长期成本,又能带来多少额外收益?答案要落在一笔更现实的账上:为了维持竞争力增加的成本,能不能被更高的商业化效率和内容效率覆盖。

这笔账对刚刚盈利的B站并不轻松。

10亿元主要属于资本开支,不会一次性进入利润表。但按照CFO樊欣在财报电话会上的口径,服务器、算力和相关研发投入,预计会增加全年约5亿元研发费用。相比2025年11.9亿元的GAAP净利润,相当于四成以上。

另一方面,截至二季度末,B站持有约243亿元现金及现金等价物、定期存款等资金。10亿元不到这一规模的5%。它不会威胁公司的现金安全,却足以影响刚刚建立起来的盈利水平。

樊欣表示,公司会削减其他运营费用,对冲部分新增成本,40%-45%的长期毛利率目标和15%-20%的营业利润率目标都没有改变。

资本市场也已经开始按照这个逻辑重新计算B站的AI投入。

年初,瑞银、摩根大通、花旗都把AI带来的广告需求和效率改善列为B站新的增长因素。二季度财报后,摩根士丹利和美银维持积极评级,却下调了目标价或盈利预测。大摩提到研发费用增加、毛利率承压,同时指出广告业务面临宏观逆风;美银同样指向研发投资增加和毛利率略低。

市场关心的问题已经从“AI有没有用”,变成了“AI带来的增量,能不能跑赢它持续增加的成本”。

 

如果每年增加的服务器、算力和研发成本,能够持续换来更高的广告转化率、更好的推荐效率和更多高质量内容,投入就能产生回报。但如果AI最终成为所有平台的标配,服务器和算力就会从“投资”变成一项“税费”——为了不掉队,年年续费。

一位长期关注互联网平台的投资人对10亿元能够创造多少增量价值持谨慎态度。

在他看来,10亿元不足以让B站获得真正的技术优势。对内容平台来说,最终拉开估值差距的仍然是用户规模、社区关系、商业化效率和产品,而不是是否部署了AI。至于B站为什么仍然需要投入,他认为还有资本市场层面的考虑:当AI已经成为科技公司的必答题,完全缺席本身就可能被视为落后。

在他看来,现在很多公司都是为了AI而AI,资本市场和公司“互相配合演出”。

吴杰茜与他的判断既有相似之处,也有分歧。两人都认为,模型本身很难成为B站独有的资产。真正的分歧在于,当同样的AI能力进入不同平台,它能否放大平台原本的差异。

在吴杰茜看来,B站的优势主要来自两端:一端是长期积累的大量C端内容创作者,另一端是黏性很高的二次元用户。AI反而可能让两项资产变得更重要。

生成式AI正在把内容生产变成更小团队的生意。现在不少AI漫剧团队只有1-5人,一个小工作室就可能做出爆款。这与B站原有的创作者结构很接近,大量UP主本来就是个人或小团队。过去需要更完整组织才能完成的内容,如今几个人就能做,B站积累多年的个人创作者供给,也因此有了新的价值。“谁拥有创作者,下一个时代的精品内容就诞生于谁。”吴杰茜说到。

而且,高黏性的用户、弹幕、评论、投币和收藏,不只是消费行为,也构成了一套持续的内容反馈机制。创作者更容易获得反馈,优质内容更容易在这里跑出来。

这是B站更愿意讲述的故事。

目前,B站有近3亿通过社区考试的正式会员,12个月留存率稳定在80%左右,每月产生约170亿次弹幕、评论、投币和收藏。陈睿把这些行为称为AI时代的“真人信号”。今年一季度,他还表示,AI会放大B站在优质内容供给和社区体验上的优势。“AI是十倍放大的历史性机遇,我们将牢牢把握”。

10亿能否撬动10倍价值仍待时间验证,但在今天的科技行业,这已经足够构成一项投资理由。毕竟,当所有公司都必须回答“你准备怎么做AI”时,不花钱本身,也会变成一种需要解释的选择。

*题图来源于哔哩哔哩官方微博。

更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App

2026-09-16 03:47:43 · 大模型,算力芯片,AI应用,融资,政策监管,字节跳动,腾讯,快手,文生图,文生视频,语音音频,搜索RAG,数字人,翻译字幕,设计创意,金融,扩散模型,微调蒸馏,传媒内容,营销广告,招聘HR,网络安全,基础设施,模型发布,产品更新,财报

智能体下线后,他们与逝去亲人的第二次“告别”

澎湃新闻
· 大模型,算力芯片,AI应用,政策监管,字节跳动,阿里巴巴,语音音频,对话助手,Agent智能体,搜索RAG,数字人,医疗健康,游戏,模型安全对齐,端侧AI,招聘HR,网络安全,产品更新,论文
AI 资讯

Habitat-GS: A High-Fidelity Navigation Simulator with Dynamic Gaussian Splatting

arXiv cs.CVarXiv:2604.12626v2 Announce Type: replace-cross Abstract: Training embodied AI agents depends critically on the visual fidelity of simulation environments and the ability to model dynamic humans. Current simulators predominantly rely on mesh-based rasterization, for which photorealistic assets are costly to author at scale, and their support for dynamic human avatars is largely constrained to mesh representations, hindering agent generalization to human-populated real-world scenarios. We present Habitat-GS, a navigation-centric embodied AI simulator extended from Habitat-Sim that integrates 3D Gaussian Splatting scene rendering and drivable gaussian avatars while maintaining full compatibility with the Habitat ecosystem. Our system implements a 3DGS renderer for real-time photorealistic rendering and supports scalable 3DGS asset import from diverse sources. For dynamic human modeling, we introduce a gaussian avatar module that enables each avatar to simultaneously serve as a photorealistic visual entity and an effective navigation obstacle, allowing agents to learn human-aware behaviors in realistic settings. Experiments on point-goal navigation demonstrate that agents trained on 3DGS scenes achieve stronger cross-domain generalization. Evaluations on avatar-aware navigation further confirm that gaussian avatars enable effective human-aware navigation, while performance benchmarks validate the system's scalability. Code is available at https://github.com/zju3dv/habitat-gs.
2026-09-15 04:00:00 · AI应用,具身智能,开源,Agent智能体,数字人,强化学习,模型评测,论文
AI 资讯

DiVA: Enabling Interactive Digital Life Simulation via Video Models

arXiv cs.CVarXiv:2609.13830v1 Announce Type: new Abstract: We present DiVA, a deeply interactive digital life simulator pioneering a new paradigm for long-term, open-ended interactive experiences within digital character worlds. DiVA's architecture pairs a Multimodal Large Language Model (MLLM) as a router with a meticulously designed stacked video pipeline for seamless, multi-turn interactions with action and audio response. To maintain continuity and avoid degradation, we model generation as a three-part coupled system: waiting video, action video, and the transitions between them. These transitions are critically handled by our Anchored Video Continuation (AVC) module, which returns the character to stable states to prevent degradation. By encoding information from the preceding action video segment, AVC ensures smooth transitions, significantly reducing camera jitter and inconsistencies common in current video transition methods. This design also enables complex pose changes (e.g., sitting to standing) typically difficult for audio-driven models. These system designs together ensure high-fidelity identity, coherence, and dynamics for extended experiences. To validate our pipeline design, we comprehensively compare our system against alternatives by replacing our core generation module with mainstream long-video, continuation, and interpolation methods. We further analyze the necessity of the three-stage design, anchor-state selection, transition naturalness, spatial grounding, and the quality-latency trade-off, and we expand the comparison to additional long-form audio-driven avatar models. Results confirm DiVA is markedly superior in maintaining long-term visual quality and realism, validating its effectiveness as a sustainable, interactive simulation.
2026-09-15 04:00:00 · 大模型,多模态,数字人,扩散模型,招聘HR,论文
AI 资讯

Beyond Scene Description: Multi-Agent Orchestration for Non-visual Access to Virtual Worlds

arXiv cs.AIarXiv:2609.14512v1 Announce Type: new Abstract: Virtual worlds now host classrooms, meetings, conferences, shops, and social venues, and nearly every interaction they expose assumes a user who can scan a three-dimensional scene, follow avatars, and read floating panels. Blind and visually impaired (BVI) users are left with assistive tools that each solve one task in isolation: naming an object, reading text, describing a scene, or planning a route. A live virtual room defeats that model: obstacles, speakers, gestures, chat, slides, and notifications arrive together, and a tool that narrates all of them trades a visual barrier for an auditory one. This paper presents MetaBlind, an architecture that distributes nonvisual access across eight specialized agents, spanning perception, navigation, social and object interaction, communication, safety and trust, memory, and personalization, and that places an Accessibility Orchestrator between those agents and the user. Agents publish candidate information into a shared accessibility context instead of speaking to the user directly. The orchestrator scores each candidate on safety relevance, goal relevance, urgency, confidence, user relevance, and estimated listening load, then releases only the items it judges relevant at that moment through speech, structured audio, or haptic output. We give the selection step a formal statement, specify the orchestration cycle as an algorithm, and define an evaluation protocol against a single-agent assistant. MetaBlind is reported at the design stage, with no prototype measurement or user study, and the protocol states which outcomes would support the design and which would refute it.
2026-09-15 04:00:00 · 算力芯片,AI应用,Google,Meta,Agent智能体,数字人,扩散模型,强化学习,招聘HR,网络安全,论文

State of Open Models: Summer 2026 Observations

HuggingFace Blog
2026-08-14T00:00:00.732Z · 大模型,算力芯片,AI应用,开源,OpenAI,Google,Meta,Microsoft,NVIDIA,阿里巴巴,DeepSeek,xAI,月之暗面,快手,文生视频,多模态,搜索RAG,数字人,Transformer,扩散模型,强化学习,微调蒸馏,模型评测,向量数据库,招聘HR,榜单评测,开发者生态

Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL

HuggingFace Blog
2026-09-10T00:00:00.744Z · 大模型,算力芯片,AI应用,自动驾驶,开源,OpenAI,Google,阿里巴巴,DeepSeek,推理思考,搜索RAG,数字人,强化学习,微调蒸馏,提示工程,招聘HR,榜单评测

担心代码被拿去训练!英伟达:限制员工使用 Claude;一汽将成广汽第二大股东!南北丰田拟合并;苹果回应「iPhone 18 Pro破发」

雷锋网要闻提示
1.突发!英伟达:限制员工使用 Claude2.重磅,一汽将成广汽第二大股东!南北丰田拟合并3.苹果回应“iPhone 18 Pro破发”:第三方渠道有自己的定价权4.比亚迪高管:燃油车没有未来5.罗永浩遭前合伙人黄斌怒喷:一辈子都在吹牛,挺可怜的6.高瓴90后合伙人严文韬或加入DeepSeek担任CFO,已发起离职流程7.程序员为用公司算力干私活:一条代码删掉89TB数据获刑

8.马斯克预热特斯拉跑车,博主李杰灵晒赛博皮卡订单称“表示怀疑”

今日头条

突发!英伟达:限制员工使用 Claude

9 月 14 日,据外媒报道,英伟达、Palantir、Booz Allen Hamilton 已开始限制内部使用 Anthropic 的 AI 模型。相关企业主要担心员工将内部代码、知识产权等敏感信息输入外部 AI 模型后,模型提供商可能接触甚至从这些数据中学习。

随着 Claude 等 AI 工具越来越多进入编程和企业办公场景,大型企业对数据如何被模型供应商保存、处理和使用的担忧也在增加。其中,英伟达目前仅在开源软件等敏感度较低的任务中使用 Anthropic 的 Fable,供应链监控等敏感内部项目则使用自家的 Nemotron 模型。(云头条)

国内资讯

重磅,一汽将成广汽第二大股东!南北丰田拟合并

9 月 14 日晚间消息,广汽集团发布公告,公司与中国第一汽车股份有限公司签署《意向协议》,筹划以发行股份方式购买一汽股份持有的某整车合资公司部分股权,并募集配套资金。经初步测算,本次交易完成后,一汽股份将成为广汽集团第二大和具有战略影响力的股东。

本次交易预计构成重大资产重组及关联交易,但不会导致广汽集团实际控制人发生变化,也不构成重组上市。广汽集团A股自9月14日开市起停牌,预计停牌时间不超过10个交易日。

另据报道,某整车合资公司指一汽丰田汽车有限公司(以下简称“一汽丰田”)。一汽股份持有一汽丰田 50% 股权,丰田汽车公司持有 45.7706%,丰田中国持有 4.2294%。

这意味着,南北丰田合并进入实质性阶段。据悉,南北丰田合并成立丰田(中国)销售公司(拟),是丰田日方力主推进的模式。新销售公司拟丰田占股 50%,一汽占股 25%,广汽占股 25%。双方旗下姐妹车型仅保留全球车,经销商渠道完全打通,可同时销售和维修所有丰田车型。(IT之家、长安街知事)

苹果回应“iPhone 18 Pro破发”:第三方渠道有自己的定价权

9月14日,某电商平台显示,256G的iPhone 18 Pro售价已经降至9099元,较苹果官网9999元的售价直降900元(官方补贴300元,店铺优惠600元)。该平台客服人员提醒:“商品进货价格会根据市场行情实时波动,所以商品是不保价的,以实时下单优惠价为准,订单成交后不退不补。”

同日,有媒体报道,某电商平台上9099元iPhone 18 Pro不支持无理由退货。该平台要求现场签收,配合快递人员拆封、激活,激活后不支持7天无理由退换货。

针对“有电商平台iPhone 18 Pro 破发”情况,苹果官方客服人员回应称:“第三方渠道有自己的定价权,他们可能有平台补贴或者店铺优惠,苹果官方不会去干预。”

该客服补充道,所有平台自营的产品在质量上应该没有任何问题,只是在售后权利方面有差异,苹果官方渠道支持14天无理由退换货,哪怕已经激活使用也没有关系,只要包装盒和手机完好无损,就可以在期间内退换货。但其他平台可能只有7天或者完全不支持退换货。消费者需要注意各平台售后规则再购买。(中新经纬)

生数科技Vidu S2突袭发布,实时视频编辑落地,布局头显空间视频

9月15日消息,在字节被传秘密推进AI视频与空间计算重磅项目之际,国内创业公司生数科技抢先亮出新品。

据X平台与行业媒体消息,生数科技将正式发布Vidu S2,距离上一代产品迭代仅69天。

新品拥有三大核心能力:S2-Editing支持视频流实时换装、换背景、切换风格,实现实时视频精修;S2-Avatar可实现720P实时数字人交互,虚拟角色能临场适配新增服饰与道具;新增空间视频功能,可将实时视频转换为适配头显的左右眼画面,抢先布局头显产品。

比亚迪高管:燃油车没有未来

9月14日消息,近日,比亚迪执行副总裁李柯在接受采访时表示:“在中国,随着比亚迪闪充技术的普及,我认为燃油车已经没有未来,这是非常明确的。”她还表示,全球其他地区的电动化转型可能需要花费几年,但是最终的结局是一样的。(第一财经)

罗永浩遭前合伙人黄斌怒喷:一辈子都在吹牛,挺可怜的

9月14日消息,近日,罗永浩的早期合伙人黄斌在微博发布长文,直接点名罗永浩,言辞激烈地对其展开批评,再度勾起两人多年前合作时期的旧矛盾。

黄斌在博文中直言,罗永浩长期夸大自身成就,甚至把自己编织的故事信以为真。他还调侃罗永浩,与其产生争执,不如继续畅想收购苹果的想法,乔布斯、库克不会回应相关言论,但自己不会一味退让。黄斌认为罗永浩并没有匹配的能力,却以企业家自居,还提到罗永浩过去在公司内部情绪激动,常有摔砸物品的行为,会对公司员工发泄怒火,但不敢和自己对峙。

黄斌还翻出了罗永浩在英语培训时期的往事。据他描述,当年老罗英语经营困难,濒临倒闭,出于昔日合作情谊,他接手了新中关大厦尚未到期的教室,帮罗永浩规避了经济损失。而就在同一时期,罗永浩外出演讲,对外宣称培训公司已经实现盈利,这件事也让黄斌耿耿于怀。

在博文结尾,黄斌评价罗永浩始终困在自我构建的幻想之中,难以认清现实,还配上“多喝热水”的表情包,表达了自己的态度。这条微博发布后,迅速收获大量互动,网友也在评论区就此展开讨论。(快科技)

高瓴90后合伙人严文韬或加入DeepSeek担任CFO,已发起离职流程

9月14日消息,据报道,高瓴创投90后合伙人严文韬,或将加入中国头部大模型公司DeepSeek出任CFO,预计将于近期正式到岗。有消息称,严文韬已经在高瓴内部发起离职流程。截至发稿,DeepSeek、高瓴方面均未就此事发布官方确认。

几天前,外媒援引知情人士消息称,DeepSeek 已聘请中信证券为其筹备境内上市,目标为上交所科创板,并计划于今年启动相关流程。按照市场传闻的时间表,2027年正式挂牌。

严文韬出生于1991年,是高瓴创投新晋合伙人,也是高瓴年轻一代投资人中的代表人物。2025年,他入选投中 Nova 新星投资人 TOP50,公开代表项目包括MiniMax、智谱、Webull、极兔、纵腾等。(让互联网飞一会)

程序员为用公司算力干私活:一条代码删掉89TB数据获刑

9月14日消息,近日,媒体公开报道了北京首例破坏人工智能模型刑事案件。据悉,涉案人员王某是一名90后算法工程师,在北京一家科技公司任职AI短剧部门,他私下和外部人员合作接私活,想要占用公司服务器算力训练外部AI模型,但服务器存储空间不足。为腾出磁盘空间存放私活项目,王某使用旧账号登录已经通知迁移数据的旧服务器集群,输入强制删除高危命令,操作完成后便直接下班离开公司。

这条删除指令在后台持续运行长达17小时,等到运维人员发现异常紧急叫停,已有89TB核心数据被清除,包含自研文生3D模型、大量训练数据集、渲染算法资产,直接导致游戏 AI 研发系统整体瘫痪。公司组织人员抢修,耗费近20天才开展数据恢复与模型重建,人力、算力合计经济损失20.4万余元,随后之后公司报警,王某被抓。

庭审阶段,王某辩称只是操作失误,但聊天记录等证据显示,他事前就在和同伙沟通私活模型调试,事后还惋惜再也得不到这么强的算力,甚至直言“下回我还敢”。因此法院不予采纳误操作的说辞,认定属于主观故意破坏计算机系统。最终法院以破坏计算机信息系统罪,判处王某有期徒刑五年十个月,同时赔偿企业20.4万余元,王某不服提出上诉,今年6月26日,二审结果出炉,维持原判。(快科技)

2026-09-15 00:23:00 · 大模型,算力芯片,自动驾驶,开源,融资,政策监管,Anthropic,NVIDIA,字节跳动,DeepSeek,xAI,智谱,代码生成,对话助手,数字人,办公效率,设计创意,金融,教育学习,游戏,扩散模型,零售电商,传媒内容,物流供应链,招聘HR,模型发布,收购并购,合作,版权诉讼,榜单评测
AI 资讯

MGAvatar: Mesh-Bound Gaussians for Head Avatar Geometry and Appearance Modeling

arXiv cs.CVarXiv:2609.12850v1 Announce Type: new Abstract: Accurate head modeling requires a stable yet expressive geometric representation. Existing Gaussian-based head avatars commonly rely on parametric templates (e.g., FLAME) for Gaussian initialization and deformation, but these templates lack personalized priors and struggle to represent structures such as hair and clothing. To address this issue, we propose MGAvatar, a Gaussian-mesh hybrid representation that jointly models geometry and appearance through two Gaussian-mesh binding modes. Specifically, we introduce vertex-bound Gaussians and constrain their learnable parameters, enabling progressive mesh deformation to represent complex head geometry, while a pose-dependent offset module accounts for non-rigid deformations. Once geometry is stabilized, MGAvatar switches to face-bound Gaussians for appearance modeling. To improve appearance consistency across novel poses and viewpoints, we introduce a view-conditioned neural color field that alleviates artifacts caused by independently optimized Gaussian colors. In addition, we design a Gaussian offset network to predict Gaussian offset maps in the observation space, providing greater flexibility for face-bound Gaussians to capture dynamic facial textures. Extensive experiments on multi-view and monocular videos show that MGAvatar outperforms existing methods in rendering quality, producing high-fidelity head avatars with rich texture details.
2026-09-14 04:00:00 · 数字人,扩散模型,招聘HR,论文
AI 资讯

HeyGen x Google Cloud: Bringing Avatar IV to TPUs

Google Developers BlogHeyGen ported their 18B+ parameter Avatar IV video generation model to Google Cloud's Trillium (v6e) TPUs via torchax and XLA, utilizing FSDP and Ulysses sequence parallelism across an eight-chip mesh. To achieve a 1.86x speedup for real-time streaming, the engineering team pipelined exposed all-to-all collectives, aligned sparse attention block sizes to eliminate mask padding, and bypassed softmax serial dependencies using a precomputed Cauchy-Schwarz upper bound. These custom Pallas kernel and compiler optimizations were deployed only after passing rigorous two-tier quality gates to guarantee byte-identical or mathematically equivalent pixel outputs.
· 算力芯片,自动驾驶,Google,数字人,Transformer

Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

HuggingFace Blog
2026-08-12T16:14:36.085Z · 算力芯片,AI应用,具身智能,开源,Google,搜索RAG,数字人,扩散模型,MoE架构,强化学习,微调蒸馏,预训练,模型评测,向量数据库,提示工程,招聘HR,收购并购,论文,开发者生态
继续滚动加载更多…