论文作者团队简介：本文第一作者周鑫，共同第一作者梁定康，均为华中科技大学博士生，导师为白翔教授。合作者包括华中科技大学陈楷锦、冯天瑞、林鸿 凯，旷视科技陈习武、丁宜康、谭飞杨和香港大学赵恒爽助理教授。 在 HunyuanVideo 上， EasyCache 在复杂场景下保持与原视频的一致外观，同时显著加速 1. 研究背景与动机 近年来，随着扩散模型（Diffusion Models）和扩散 Transformer（DiT）在视频生成领域的广泛应用，AI 合成视频的质量和连贯性有了飞跃式提升。像 OpenAI Sora、HunyuanVideo、Wan2.1 等大模型，已经能够生成结构清晰、细节丰富且高度连贯的长视频内容，为数字内容创作、虚拟世界和多媒体娱乐带来了巨大变 革。 但与此同时，推理慢、算力消耗高的问题也日益突出。以 HunyuanVideo 为例，生成一个 5 秒、720P 分辨率的视频，单次推理在单张 H20 上需要 2 小时。这种高 昂的资源代价，极大限制了扩散视频生成技术在实时互动、移动端和大规模生产场景的应用落地。 造成这一瓶颈的核心原因，是扩散模型在生成过程中需要多次迭代去噪，每一步都要进 ...

Core Insights - ZyVersa Therapeutics is advancing its Cholesterol Efflux Mediator VAR 200, which has received FDA authorization for Emergency Compassionate Use in a patient with ApoCII amyloidosis [2][3] - The company is conducting a Phase 2a clinical trial for diabetic kidney disease (DKD) to gather data that will support the ongoing development of VAR 200 [3][4] - ZyVersa has entered into a warrant inducement agreement with an institutional investor, raising $2 million through the exercise of Series A-2 and A-3 warrants at a reduced price [7] Company Developments - VAR 200 is an injectable drug in Phase 2 development aimed at reducing renal lipid accumulation, which is critical for kidney health [4][5] - Preclinical studies have shown that VAR 200 can lower cholesterol and lipid levels, protect against renal injury, and improve proteinuria in animal models [5][6] - The company is exploring additional indications for VAR 200, including Alport Syndrome and diabetic kidney disease, as part of its indication expansion strategy [6] Market Performance - ZyVersa's stock (ZVSA) experienced a significant increase of 56%, reaching $1.03, with a trading volume of 160.9 million shares compared to the average of 4.03 million [1][7]

Core Viewpoint - ZyVersa Therapeutics, Inc. has entered into a warrant inducement agreement with an institutional investor, which includes the immediate exercise of existing warrants and the issuance of new warrants at a reduced exercise price of $0.67, generating approximately $2.0 million in gross cash proceeds for working capital and corporate purposes [1][2]. Group 1 - The agreement involves the immediate exercise of Series A-2 Warrants for up to 957,200 shares and Series A-3 Warrants for up to 2,105,265 shares at an exercise price of $0.67 [1]. - The investor will receive new Series A-4 Warrants to purchase up to 6,124,930 shares, which will be exercisable upon stockholder approval and will expire five years from that date [2]. - The transaction is expected to close on or about July 9, 2025, pending customary closing conditions [2]. Group 2 - A.G.P./Alliance Global Partners served as the exclusive financial advisor for this transaction [3]. - The new warrants were offered in a private placement under an exemption from the registration requirements of the Securities Act, and the company plans to file a registration statement with the SEC for the resale of common stock upon exercise of the new warrants [4]. Group 3 - ZyVersa Therapeutics is a clinical stage specialty biopharmaceutical company focused on developing first-in-class drugs for renal and inflammatory diseases, utilizing proprietary technologies such as Cholesterol Efflux Mediator™ VAR 200 and Inflammasome ASC Inhibitor IC 100 [6].

以下文章来源于AGI接口 ，作者陈伊凡 AGI接口 . AI卷起的财富风暴。 出品｜虎嗅科技组 作者｜陈伊凡、孙晓晨 编辑｜苗正卿 头图｜《硅谷》剧照 "AI 原生 100" 是虎嗅科技组推出针对 AI 原生创新栏目，这是本系列的第「 03 」篇文章。 当同行还在卷模型参数时，两个辍学生把50个开源AI模型"塞"进了一个画布——点几下鼠标、随手 涂鸦，40毫秒内图像跃然眼前。 Krea，一家AI设计公司，在成立16个月后，年经常性收入（ARR）冲 到800万美元，收入翻了20 倍，用户超2000万，成为入选硅谷最大加速器YC的项目中 "ARR增速最快的公司之一"。同时，Krea 也入选了硅谷顶尖风投机构A16Z GenAI 100的榜单，该榜单收录了当下最有价值的100个AI应用。 资本市场也给了Krea更高的想象空间。根据公开资料，虎嗅测算了Krea的PS值，PS值通俗而言就是 投资者愿意为每1元销售收入支付多少钱。PS值高意味着高市场预期。根据Krea的800万美元年收入 和5亿美元的估值，其PS值为62.5倍，Midjourney的估值为100亿美元，2024年年收入为3亿美元，估 值为33.3倍。 虎 ...

本论文共同第一作者闫泽轩和马跃分别是上海交通大学人工智能学院 2025级研究生，以及香港科技大学 2024级博士生。目前在上海交通大学EPIC Lab进行科研实习，接受张林峰助理教授指导，研究方向是 高效模型和AIGC。 本文主要介绍张林峰教授的团队的最新论文： EEdit ⚡ : Rethinking the Spatial and Temporal Redundancy for Efficient Image Editing。 这是首个用于加速匹配流模型上兼容多种引导方案的图像编辑框架。该框架速度提升显著，较原始工作 流可加速2.4倍；并且输入引导条件灵活，支持包括参考图像引导，拖拽区域引导，提示词引导的多种 编辑任务；该框架采用免训练的加速算法，无需微调和蒸馏。 该论文已经入选ICCV 2025。 论文链接： https://arxiv.org/pdf/2503.10270 论文已开源： https://github.com/yuriYanZeXuan/EEdit 最近，基于流匹配（ Flow Matching）的扩散模型训练方式逐渐成为扩散模型的热点，以其优雅简洁 的数学形式和较短时间步的生成能力吸引 ...

本周核心趋势 2025年7月2日 1、「重磅事件」 Meta疯狂挖角OpenAI华人AI大牛：硅谷上演人才争夺战 Meta创始人扎克伯格以人均 1亿美元年薪+无限制算力 挖走8名OpenAI核心研究员，其中7人为华人顶尖 人才！试图组建50人"超级智能团队"对抗OpenAI。 本周AI商业焦点必读 （2025.6.24-7.3） 巨头抢人，开源崛起： Meta天价挖角OpenAI，暴露巨头人才垄断危机，开源模型将加速普及，小 公司机会变多！ AI智能体别光吹牛： Gartner警告40%项目要黄，烧钱不实用的AI智能体（Agent）会死掉，能真干 活的才能活下来！ AI赋能，改造老行当： 美团用数据帮开店，小米做AI眼镜，AI正深入改造餐饮、硬件等传统行业， 落地为王！ 小模型猛，开源真香： 智谱、快手的小模型性能炸裂，黑森林开源叫板巨头，好用又便宜的开源小 模型是创业新利器！ 这场争夺战暴露了AI行业顶尖人才垄断化趋势，Meta的Llama 4模型失利加速了人才掠夺，而OpenAI面 临GPT-5研发窗口期被压缩的危机。未来，随着xAI等新势力加入战局（如马斯克团队同期发布Grok 4），全球AI人才流动或 ...

跟海外功能一样的同时，对中文内容生产更加友好 ，在这段时间还加上了一致性无敌的 FLUX Koontext 模型和可以直接生成语音音效的视频模型，更加强大了。 同时价格也是 相当划算，首月算力五折，无需要请码！ 这次我也整理了我专门为这种批量设计 Agent 写的一些提示词： 大家好，我是歸藏（guizang），今天带来 Lovart 国内版星流 Agent 的介绍和提示词合集。 前段时间 Lovart 的推出真的引起了轰动，而且效果也很好，基本是我觉得这波 Agent 热潮中完成度和稳定性最高的垂类 Agent 产品了。我也写过详细的流程拆解 文章《 不看这藏师傅篇深度拆解，你永远不知道Lovart AI有多可怕 》。 现在 Lovart 的国内版本 星流 Agent（ https://www.xingliu.art/ ） 正式上线。 首先是传统艺能，图像海报的批量生成。 我在海外版本的时候探索了很多适合设计 Agent 用的同体系批量创意图像生成提示词，刚好可以拿来用。 先来广受好评的 Q 版国风戏曲塔罗牌设计。 这个提示词除了考察模型对于画风一致性的表现，也可能考量模型对于世界知识的理解，他需要了解每张 ...

Black Forest官方测试报告显示，该模型在人类偏好评估、指令编辑等多项关键指标优于OpenAI最新发 布的GPT-image-1，标志着开源模型在高精度图像编辑领域取得新进展。 FLUX.1-Kontext的技术架构由自然语言解析、图像生成和多模态融合三个关键模块构成。 自然语言解析层采用改进型Transformer架构，配置8层自注意力机制，能对用户指令做深度语义拆分。 比如面对"将画面左侧咖啡杯替换为青花瓷杯，杯内咖啡表面添加拉花图案"的指令，系统可精准识别出 对象替换、材质变更、细节添加等子任务，并分配相应权重。 图像生成引擎基于改进版扩散模型（DPM-Solver++）构建，创新引入动态噪声调度机制。 作者：周源/华尔街见闻 在AI绘画领域竞争白热化的当下，开源与闭源模型的博弈持续深化。 6月底，知名开源平台Black Forest推出文生图模型FLUX.1-Kontext开发者版本，凭借"自然语言指令实现 图像编辑"的核心功能，迅速成为行业焦点。 该机制可依据指令复杂程度自动调整去噪迭代次数：处理"将天空改为黄昏色调"等简单指令时，20步内 即可完成；面对"将人物服装添加复古刺绣纹样"等复杂 ...

Group 1: Apple Developments - Apple is reportedly developing multiple new wearable products to address declining market performance in its wearable segment, including a new Apple Watch SE and Apple Watch Ultra 3, expected to feature significant updates such as satellite connectivity independent of the iPhone [4][5] - There are indications that Apple may launch a smart ring product that would integrate with other devices, targeting users who may not wish to purchase an Apple Watch [5][6] - The smart ring is seen as more suitable for sleep monitoring and offers longer battery life due to the absence of a display, potentially allowing Apple to capture a different segment of the market [6] Group 2: Tesla's Autonomous Delivery - Tesla has completed its first fully autonomous vehicle delivery without a driver or remote control, achieving a top speed of 115 km/h [10] - The company is also testing its Robotaxi service, which utilizes ten years of developed autonomous driving technology, although there are concerns about safety following reports of incidents during initial testing [10][11] Group 3: Meta and OpenAI Talent Acquisition - Meta has successfully recruited four researchers from OpenAI, who were involved in key projects like ChatGPT and GPT-4, to join its new "superintelligence" team [12][13] - OpenAI's leadership has expressed strong concerns over this talent acquisition, indicating a competitive response to retain their workforce [14] Group 4: Neuralink's Future Plans - Neuralink has implanted brain-machine interfaces in seven volunteers, allowing them to control games and mechanical arms through brain signals, with a long-term vision of creating a universal brain interface [15] - The company has outlined a roadmap for the next few years, aiming to restore vision for the blind and enhance treatment for neurological disorders [15] Group 5: Xiaomi's Vehicle Response - Xiaomi has addressed concerns regarding a reported fire incident involving the YU7 vehicle's brake pads during track testing, clarifying that the incident was due to high temperatures and did not affect the overall braking system [17][18] Group 6: Trump's Phone Production - The CEO of Purism has stated that the T1 Phone, launched by Trump's mobile company, is manufactured in China, contradicting earlier claims of being "American-made" [19][20] Group 7: Microsoft's AI Vision - Microsoft's CEO emphasized the underestimated potential of AI in improving the lives of everyday people, advocating for broader access and application of AI technologies [21][23] - He also discussed the future integration of AI tools, suggesting a transformative moment akin to the introduction of web browsers [23] Group 8: Huawei's New Patent - Huawei has unveiled a new patent for a three-fold smartphone design, which differs from its previous models and is expected to be part of its upcoming product lineup [28][29] Group 9: Qualcomm and Samsung Collaboration - Samsung is set to manufacture part of Qualcomm's Snapdragon 8 Elite Gen 2 processors using its 2nm process technology, marking a shift from exclusive reliance on TSMC [30]

鹭羽 发自 凹非寺 量子位 | 公众号 QbitAI 图像模型开源还得是FLUX！ Black Forest Labs刚刚宣布开源旗舰图像模型 FLUX.1 Kontext[dev] ，专为图像编辑打造，还能直接在消费级芯片上运行。 只有小小的 12B ，更少的参数，更快的推理，性能更是媲美 GPT-image-1 等一众闭源模型。 现在FLUX.1 Kontext[dev]可以让小狗迅速离开画面，为小老鼠戴上胡须，添加文字、修改背景也不在话下。 或者多次输入指令， 直到让小哥成为酒吧里最靓的崽（bushi） ，直到让画面符合咱们需求。 具体来说，FLUX.1 Kontext[dev]的主要特点有： 网友们也立马上手试玩，制作了一个旅行的CPU青蛙？ 1. 可以根据编辑指令直接更改现有图像，以及进行精确的 本地和全局编辑 。 2. 不用做任何微调，就能 直接引用 里面的人物角色、风格样式和物品元素。 3. 允许用户通过 多次连续编辑 优化图像,同时将视觉漂移降到最低。 4. 专门为NVIDIA Blackwell进行了 权重优化 。 旅行必备的墨镜，还有抗寒的帅气红色毛衣也要准备妥当。 （蛙蛙：出片，我势在必 ...