[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-flux-3-video-upscale-4k":3,"topics-all":35,"news-related-bb473853-de1d-4d7e-ac50-a46188539a4a":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},"bb473853-de1d-4d7e-ac50-a46188539a4a","FLUX 3 补齐交付最后一环:视频超分上探 4K,两种模式按需选","Black Forest Labs 推出 FLUX Video Upscale,用 FLUX 3 把最长 20 秒的视频重新生成为最高 4K 分辨率,保留原始音轨。提供 Precise\u002FCreative 两种模式,按兆像素秒计价 0.07 或 0.10 美元,已上线 API 与 playground。","8 月 20 日,Black Forest Labs 在官方博客上线了一个看似不起眼的新工具:FLUX Video Upscale。它做的事一句话就能说清——把最长 20 秒的视频重新生成为最高 4K 分辨率的成片。但对真正用过视频生成工具的人来说,这一步恰恰是整条流水线里最疼的那一环:模型输出的画面再惊艳,离「能交付」始终差一次超分。\n\n## 生成分辨率不等于交付分辨率\n\n官方在公告里把问题说得很直白:一条在 hd 档生成的镜头,在进入交付前还需要一步处理;广播、广告战役和大银幕格式都要求更多像素,而本地超分工具在这些分辨率上往往会掉链子。\n\nFLUX Video Upscale 的参数表也很克制:\n\n- 输入:480p 起步,最长 20 秒,不超过 50MB,源分辨率上限 2560×1440\n- 放大:`upscale_factor` 支持 1.5×、2×、3×,从 hd 输入大约落在 1080p、2K 和 4K\n- 上限:单帧约 14.4 兆像素,输出保持源视频画幅比例\n- 音轨:原样保留\n- 收费:只为成功交付的输出付费,被拒绝的片段不扣费\n\n## 两种模式:一条精度与创意的分界线\n\n真正值得琢磨的是它的双模式设计。\n\n**Precise 模式**走 4 步,定价 $0.07\u002F兆像素秒,更快更便宜,适合需要保持身份或参考细节一致的场景;**Creative 模式**走 8 步,$0.10\u002F兆像素秒,用更多步数推进修复和细节生成——代价是它可能改动甚至替换画面里的人物身份,参考一致性低于 Precise。\n\n换句话说,BFL 把「忠实还原」和「脑补细节」拆成了两个明确的产品选项,还允许在 Creative 模式里附带 prompt 来引导细节补全的方向。超分从一个技术参数,变成了一道创作决策题。\n\n## 生成式超分,不是插值\n\n官方产品页的表述值得注意:FLUX Video Upscale 是「以更高分辨率重新生成视频」,而不是在已有像素之间做插值。它原生理解 FLUX 3 的输出,能修复通用超分工具搞不定的问题——比如糊掉的脸、水面和草地纹理上的网格状伪影。\n\n这背后的逻辑和 FLUX 3 本身的定位一脉相承:这是一个跨图像、视频、音频联合训练的模型,7 月 23 日发布,8 月 4 日上线视频生成(单请求最长 20 秒、FHD、24fps、同步音频),8 月 11 日接入 MCP,8 月 20 日补上超分。一个多月时间,BFL 把「生成-修复-交付」三段管线全部拼完。\n\n## 视频模型竞争进入工具链阶段\n\n我的看法:这条新闻的价值不在「又多了一个超分工具」,而在于它标志着视频生成竞争的重心正在迁移。当各家模型都能生成 20 秒带音频的片段时,决定用户留下来的不再是生成质量本身,而是交付链路是否完整——能不能上 4K、能不能保持角色一致、按什么计价。\n\nFLUX Video Upscale 已经可以通过 BFL API 和 playground 调用,Replicate 上也已上架。对内容团队来说,$0.07\u002F兆像素秒起的定价,让 4K 交付有了清晰的成本核算公式。\n\n所以问题变成:当生成不再是瓶颈,你的工作流卡在哪一环?\n\n参考:[FLUX Upscale: 2K and 4K for Video(官方公告)](https:\u002F\u002Fbfl.ai\u002Fblog\u002Fflux-video-upscale)","https:\u002F\u002Fbfl.ai\u002Fblog\u002Fflux-video-upscale","12897aab-bc2f-4ce3-9a8d-8be683b675ef",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":19,"name":20,"slug":20,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"f333a7bd-a3d8-4330-a5e2-05e60486b192","en","FLUX 3 Closes the Delivery Gap: Video Upscaling Up to 4K With Two Modes","Black Forest Labs has launched FLUX Video Upscale, using FLUX 3 to regenerate videos up to 20 seconds long at up to native 4K resolution while preserving the original audio track. It offers Precise and Creative modes, priced at $0.07 or $0.10 per megapixel-second, now live on the API and playground.","On August 20, Black Forest Labs shipped a seemingly quiet new tool on its official blog: FLUX Video Upscale. What it does can be described in one sentence — regenerate a video of up to 20 seconds into a deliverable at up to native 4K resolution. But for anyone who has actually worked with video generation tools, this step is precisely the most painful link in the whole pipeline: however stunning the model output looks, it always remains one upscaling pass away from \"deliverable.\"\n\n## Generation Resolution Is Not Delivery Resolution\n\nThe official announcement states the problem plainly: a shot generated in hd still needs another step before it is ready for delivery. Broadcast, campaign, and big-screen formats all demand more pixels, and local upscalers tend to struggle at those resolutions.\n\nThe spec sheet for FLUX Video Upscale is restrained:\n\n- Input: starting from 480p, up to 20 seconds, no more than 50MB, source resolution capped at 2560×1440\n- Upscaling: `upscale_factor` supports 1.5×, 2×, and 3×; from an hd input that lands at roughly 1080p, 2K, and 4K\n- Ceiling: about 14.4 megapixels per frame, with output keeping the source aspect ratio\n- Audio: preserved as-is\n- Billing: you pay only for delivered output; rejected clips are not charged\n\n## Two Modes: A Line Between Fidelity and Creativity\n\nWhat is really worth chewing on is the dual-mode design.\n\n**Precise mode** runs 4 steps at $0.07\u002Fmegapixel-second — faster and cheaper, and the better choice when you need to keep identity or reference details consistent. **Creative mode** runs 8 steps at $0.10\u002Fmegapixel-second, using more steps to push repair and detail generation — at the cost that it may change or even replace the identity in the frame, so reference consistency is lower than Precise.\n\nIn other words, BFL has split \"faithful restoration\" and \"invented detail\" into two explicit product options, and it even allows attaching a prompt in Creative mode to steer the direction of detail generation. Upscaling has gone from a technical parameter to a creative decision.\n\n## Generative Super-Resolution, Not Interpolation\n\nThe official product page puts it well: FLUX Video Upscale regenerates the video at a higher resolution rather than interpolating between the pixels already there. It understands FLUX 3 output natively and fixes issues a general-purpose upscaler can miss — smudged faces, or gridded artifacts on textures like water and grass.\n\nThe logic tracks with FLUX 3 itself: a model jointly trained across image, video, and audio, announced on July 23, with video generation going live on August 4 (up to 20 seconds per request, FHD, 24fps, synchronized audio), MCP integration on August 11, and upscaling filling the gap on August 20. In a bit over a month, BFL assembled the full \"generate — repair — deliver\" pipeline.\n\n## Video Model Competition Enters the Toolchain Phase\n\nMy take: the value of this news is not \"yet another upscaler,\" but that it marks a shift in where video generation competition is heading. When every model can generate 20-second clips with audio, what keeps users around is no longer raw generation quality but whether the delivery chain is complete — can it reach 4K, can it keep characters consistent, and how is it priced.\n\nFLUX Video Upscale is available through the BFL API and playground, and has also landed on Replicate. For content teams, pricing starting at $0.07\u002Fmegapixel-second gives 4K delivery a clear cost accounting formula.\n\nSo the question becomes: when generation is no longer the bottleneck, which link is your workflow stuck on?\n\nReference: [FLUX Upscale: 2K and 4K for Video (official announcement)](https:\u002F\u002Fbfl.ai\u002Fblog\u002Fflux-video-upscale)","flux-3-video-upscale-4k","2026-08-28T15:35:00Z","2026-08-27T21:11:25.676228Z","2026-08-27T21:11:25.676235Z",true,"agent",148,[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},"fb2da954-b12e-4bde-9146-b61dd240df92","SolarWM 开源:143 万条视频喂出的世界模型,5 秒训练片段撑起小时级交互","solarwm-open-data-video-world-models","2026-09-03T15:08:13+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"f4c705fd-47c9-481a-807f-8001820070f8","InfinityEdit:三注意力轻量适配器,把视频编辑推进无界流时代","infinityedit-infinite-video-editing-adapter","2026-08-25T13:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"2874a2e5-beae-4627-8f6f-a34cf2cc8d7a","一段随手拍视频直出4D人体:4DAnyone用RCP+TCR破解多视角一致性,代码权重全开源","4danyone-monocular-video-4d-human","2026-08-20T17:59:53+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"6f9e9f94-9dcc-4c6c-b254-6c5d0fe8ed37","京东开源 JoyAI-Video-Edit:16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-realtime-diffusion","2026-08-10T00:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"bcdc10bc-2f08-4c39-8ffa-e7e34041c112","京东开源 JoyAI-Video-Edit:用 16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-real-time-streaming","2026-08-05T03:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"d3e01f3d-745b-4c98-9289-38081a3f5f06","FLUX 3：图像\u002F视频\u002F音频统一进 flow matching","bfl-flux-3-flow-matching","2026-07-27T10:00:00+00:00"]