[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openshot-40-local-ai-object-mask":3,"topics-all":35,"news-related-265ac7bc-a2a2-4b46-9434-c11a242099d1":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"265ac7bc-a2a2-4b46-9434-c11a242099d1","OpenShot 4.0 把 YOLO、EfficientSAM 塞进桌面编辑器:本地模型终于不用订阅了","OpenShot 4.0 释出,核心更新包括新 Color View、Recording View、内置本地 ML 模型驱动的 Object Mask(基于 YOLO\u002FEfficientSAM\u002FCutie ONNX),以及 libopenshot 1.0。整套 AI 蒙版与目标检测完全跑在用户本机,不依赖云端订阅。","视频编辑软件给人的印象通常是\"云端订阅 + 一键 AI\"。OpenShot 4.0 在 8 月 30 日释出,选择把这条路径反过来走:把 YOLO、EfficientSAM、Cutie 这几个原本跑在服务器上的模型,改造成 ONNX 格式直接塞进桌面客户端,全程在本机执行,不需要账号、不上传素材、不接 API。\n\n## 把 AI 从云端搬到桌面\n\nOpenShot 新加的 Object Mask 效果和 Object Detection 工作流,是这次 4.0 最值得 AI 从业者关注的点。官方在 [openshot-onnx](https:\u002F\u002Fgithub.com\u002FOpenShot\u002Fopenshot-onnx) 仓库里开源了转换好的模型集合,包括 YOLOv5、EfficientSAM、Cutie 这三类,统一转成 ONNX 格式,由 libopenshot 和 OpenCV 在本地调用。\n\nObject Mask 的用法很简单:在视频某一帧上点几个正样本点和负样本点,OpenShot 先给出分割预览,再把这张蒙版沿时间轴\"跟踪\"下去。目标变形或者被遮挡的时候,可以再补点 prompt。生成的蒙版可以直接作为 Blur、Pixelate、Color Grade 等效果的输入,既能只处理主体,也能反相处理背景。\n\nObject Detection 也升级了:支持可下载的 YOLOv5 ONNX 模型、模型校验、分割蒙版,以及更细的对象框\u002F标签\u002F变换控制。整个特性被官方明确表述为 \"Local models, local footage\"——素材不上传,功能不依赖付费云服务。\n\n## 不只是 AI:Color View 和 Recording View\n\n4.0 同时给非 AI 用户准备了两件大事:\n\n- **新 Color View**:把调色、曲线、LUT、波形监视器、RGB Parade、Vectorscope 整合到同一视图里,带肤色参考线和区域采样工具。所有调色参数都关键帧化,适合做动画色彩过渡。配套新增的 Color Grade effect 涵盖曝光\u002F对比\u002F色温\u002F色调\u002F高光\u002F阴影\u002F饱和度\u002F自然饱和度。\n- **新 Recording View**:在编辑器里直接录麦克风、桌面、摄像头、系统音频,每路信号单独成轨,可以单独裁剪、加效果而不互相影响。屏幕+摄像头组合会自动生成圆角画中画。支持 Windows\u002FmacOS\u002FLinux X11\u002FLinux Wayland(通过 PipeWire)。\n\n## 性能翻倍\n\nOpenShot 团队给出 4.0 与 3.5.1 的基准对比:Blur 工作负载快 **61.8%**,Sharpen 快 **12.8%**,时间轴渲染快 **3.4%**,带变换的时间轴渲染快 **5.1%**。Blur 的提升尤其大,源于底层 libopenshot 重写了图像处理路径。Scopes、Color Grade、Film Grain、Audio Visualization 也都顺手做了性能优化,填充波形和颜色分析路径简化了不少。\n\n## 平台层面的现代化\n\n底层的 libopenshot 和 libopenshot-audio 同步迈入 **1.0.0**,引入了新效果、实时捕获 reader、色彩分析、本地目标蒙版和 Qt 6 支持。代码层为 Android 做了大量基础工作(文件访问、渲染、Qt 绑定、大 ARM64 内存寻址),但官方明确说 4.0 不会发布 Android 版本,这是为后续实验铺路。\n\n## 一点想法\n\n视频编辑的\"AI 化\"过去两年被云端订阅叙事主导——剪映、CapCut Pro、Adobe Firefly 之类,功能上确实强,但用户要么按月付钱,要么把素材交给别人。这次 OpenShot 4.0 的解法很克制:把现成的开源视觉模型(YOLO 检测、EfficientSAM 分割、Cutie 跟踪)转成 ONNX,跟桌面引擎解耦,门槛是用户自己的显卡和 CPU,而不是订阅费。\n\n它不会一夜之间颠覆专业剪辑市场,但对独立创作者、教育场景、本地化素材处理来说,这条路证明了一件事:**消费级显卡 + 开放权重模型 + ONNX 运行时**,已经足够覆盖相当一部分\"AI 蒙版 \u002F 目标跟踪\"工作流。云端不再是唯一答案。\n\n更多细节见 [OpenShot 4.0 官方公告](https:\u002F\u002Fwww.openshot.org\u002Fblog\u002F2026\u002F08\u002F30\u002Fopenshot-40-record-edit-color-like-never-before\u002F)。","https:\u002F\u002Fwww.openshot.org\u002Fblog\u002F2026\u002F08\u002F30\u002Fopenshot-40-record-edit-color-like-never-before\u002F","0001806e-f554-4459-bb4f-9f791a430291",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"3cdd2152-064a-49a2-8deb-b2bc97929956","en","OpenShot 4.0 puts YOLO and EfficientSAM in the desktop editor: local models, no subscription needed","OpenShot 4.0 ships with a Color View, a Recording View, and a local-only Object Mask powered by YOLO, EfficientSAM, and Cutie ONNX models. The whole AI mask and detection pipeline runs on the user's machine — no cloud, no subscription, no footage upload. libopenshot also reaches 1.0.","When people hear \"AI video editor,\" they usually picture cloud subscriptions and one-click upscaling. OpenShot 4.0, released on August 30, runs the other direction: it ships YOLO, EfficientSAM, and Cutie — models that normally live on a server — as ONNX files inside the desktop client. Inference happens locally, with no account, no footage upload, and no API calls.\n\n## Bringing AI from the cloud to the desktop\n\nThe most AI-relevant pieces of 4.0 are the new **Object Mask** effect and the upgraded **Object Detection** workflow. The team open-sourced the converted models in the [openshot-onnx](https:\u002F\u002Fgithub.com\u002FOpenShot\u002Fopenshot-onnx) repo, covering YOLOv5, EfficientSAM, and Cutie — all converted to ONNX so libopenshot and OpenCV can run them on the user's machine.\n\nObject Mask is straightforward: drop a few positive and negative points on a frame, OpenShot generates a segmentation preview, then tracks that mask across the timeline. When the subject changes shape or gets occluded, users can add more points. The finished mask plugs directly into Blur, Pixelate, Color Grade, and other effects — it can act on the subject, be inverted to act on the background, or be combined for more advanced composites.\n\nObject Detection also received a major upgrade: downloadable YOLOv5 ONNX models, model validation, segmentation masks, and finer controls over boxes, labels, and tracked-object transforms. The team's wording is unambiguous — \"Local models, local footage.\" The footage never leaves the machine, and the feature does not require a paid cloud subscription.\n\n## Not just AI: Color View and Recording View\n\n4.0 ships two big additions for non-AI users:\n\n- **New Color View** unifies color wheels, curves, LUTs, and live scopes (Luma Waveform, Histogram, RGB Parade, Vectorscope) in one workspace, with a skin-tone reference line and a region probe. Every parameter is keyframable, so a stylized color transition can be animated on the timeline. The accompanying **Color Grade** effect bundles exposure, contrast, temperature, tint, highlights, shadows, saturation, and vibrance controls.\n- **New Recording View** lets users capture microphone, screen, webcam, and system audio directly inside the editor. Each source lands as its own clip on its own track, so you can trim or apply effects to one without touching the others. Combined screen + webcam captures auto-generate a rounded picture-in-picture layout. Windows, macOS, Linux X11, and Linux Wayland (via PipeWire) are all supported.\n\n## Performance doubles in the right places\n\nThe team's 4.0-vs-3.5.1 benchmarks show: **Blur** runs **61.8% faster**, Sharpen **12.8% faster**, timeline rendering **3.4% faster**, and timeline-with-transforms **5.1% faster**. The Blur jump comes from a rewrite of the image processing path inside libopenshot. Scopes, Color Grade, Film Grain, and Audio Visualization were also reworked; filled waveform drawing and color analysis both got faster shortcuts.\n\n## A modern foundation\n\nBoth libopenshot and libopenshot-audio reach **1.0.0** in this release, picking up new effects, live-capture readers, color analysis, local object masking, and Qt 6 support. The codebase also lays groundwork for Android — file access, rendering, Qt bindings, large ARM64 memory addresses — but the team is explicit that 4.0 is **not** an Android release. It's preparation, not launch.\n\n## A short take\n\nThe \"AI-ification\" of video editing over the past two years has mostly been a cloud-subscription story — CapCut Pro, Adobe Firefly, and similar tools are powerful, but users either pay monthly or hand their footage to a third party. OpenShot 4.0 takes the opposite bet: take well-known open-weight vision models (YOLO for detection, EfficientSAM for segmentation, Cutie for tracking), convert them to ONNX, and decouple them from any cloud runtime. The cost shifts from a subscription fee to the user's own GPU and CPU.\n\nIt won't upend the pro NLE market overnight, but for independent creators, classrooms, and anyone handling sensitive footage, it shows a workable path: **consumer hardware + open-weight models + the ONNX runtime** is enough for a real chunk of the AI-mask-and-track workload. The cloud is no longer the only answer.\n\nMore detail in the [official OpenShot 4.0 announcement](https:\u002F\u002Fwww.openshot.org\u002Fblog\u002F2026\u002F08\u002F30\u002Fopenshot-40-record-edit-color-like-never-before\u002F).","openshot-40-local-ai-object-mask","2026-09-02T03:00:00Z","2026-09-02T07:03:56.000957Z","2026-09-02T07:03:56.000964Z",true,"agent",108,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"70ea74fc-77ed-49a2-8498-f24edd822970","AI 工具把 Linux 内核挖出 2000 个 CVE,维护者快扛不住了","linux-kernel-ai-cve-overflow","2026-09-09T05:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","one-shot-opd-single-query-distillation","2026-09-05T21:07:11+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"4a89fe5a-8703-49e5-b083-079cbda0fa2a","蒸馏也有副作用:中间训练期上KD,推理上涨、事实记忆反而变慢","switch-distillation-midtraining-kd","2026-09-02T17:10:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"454286bd-cb8e-462e-8b33-1b4c77b27262","470M 语音模型 1 秒转写 3.5 小时:IBM 把 ASR 里的语言模型砍掉了","granite-speech-5-turbo-ctc-470m","2026-08-31T15:10:00+00:00"]