[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kandinsky-6-0-video-audio-generation-open-source":3,"topics-all":38,"news-related-ab13d62e-1d32-4af6-9d11-ebdb601ff274":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"ab13d62e-1d32-4af6-9d11-ebdb601ff274","Kandinsky 6.0 开源:29B 音视频一体生成","Kandinsky Lab 发布并开源 Kandinsky 6.0 Video:3B Lite 与 29B Pro 双模型直接生成 5 秒带 44kHz 同步音频的视频,支持唇形同步。双流 CrossDiT 架构加 RL 后训练把语音词错率降低 47%,权重、代码与 diffusers 集成以 MIT 协议放出。","视频生成模型长期有个尴尬的默认设置:片子是哑的。配音、音效、环境声要靠 TTS 工具和剪辑软件后期拼上去,对口型更是人工活。10 月 4 日,Kandinsky Lab 放出论文《Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation》([arXiv:2610.05608](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.05608)),两天后的 10 月 6 日直接把权重、代码和 diffusers 集成以 MIT 协议全部开源,把\"音视频一体生成\"从闭源旗舰的卖点变成了开源社区的公共能力。\n\n## 两个尺寸,一次成片\n\n这次发布是一对模型:Lite 版 3B 参数,Pro 版 29B 参数。两个模型都直接输出 5 秒视频片段,自带 44kHz 同步音频和唇形同步,支持文生音视频(T2AV)和图生音视频两种模式;再挂一个 1.4B 参数的超分扩散模型,可以把成片拉到 1920×1080。团队规模不小,论文署名作者超过 80 人。\n\n## 双流 CrossDiT:视频流带音频流\n\n架构上最值得看的是\"双流\"设计:视频流承接 Kandinsky 5.0 的视频生成能力,音频流从零开始,先在大规模纯音频语料上预训练,再与视频流在成对音视频数据上联合训练;两条流之间用双向交叉注意力做时间和语义对齐,同时保持单模态保真度不退化。预训练之后是监督微调、强化学习后训练和蒸馏三步。官方数据显示,RL 后训练把 Pro 版的语音词错率(WER)降低了 47%——这是本次发布所有数字里最硬的一个。\n\n官方自建评测 VABench 上,Pro 版在语音质量、音频美感、音画对齐、唇形同步、视觉真实感等项目上优于 LTX 2.5 与自家 Lite;人工盲评中 Pro 对上一代 Kandinsky 5.0 Pro 是\"明显胜出\",对 LTX 2.5 在画面质量、运动真实感、伪影控制等维度占优,语音质量优势具统计显著性。注意这些是官方口径的评测,结论以论文原文为准。\n\n## 消费级显卡能跑,生态当天就位\n\nREADME 给出的实测数据很实在:非蒸馏版 Lite 在 RTX 4090 上一段 5 秒成片约 437 秒,H100 约 239 秒;Pro 版 Full-HD 在 H100 上约 402 秒,RTX 5090 用 block offload 约 854 秒——慢,但消费级显卡确实能跑。开源当天(10 月 6 日)生态同步就位:ComfyUI 节点包、vLLM-omni 支持、Hugging Face 在线 Demo 一同上线,GitHub 仓库一天内拿到 120+ star,HF 上已有 8 个关联模型仓库。\n\n## 所以呢\n\n音视频一体生成此前基本是闭源旗舰的领地,开源方案多靠\"视频模型 + TTS\"管道拼接,对齐和唇形同步是老大难。Kandinsky 6.0 用 MIT 协议把整套东西——双基础模型、超分模型、蒸馏版——一次放全,对本地部署和二次开发相当友好。真正的悬念在长视频:5 秒是当前上限,把同步生成扩展到分钟级,才是下一场硬仗。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.05608","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"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},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2ef05a99-e67a-4c98-81de-ecfc9c3ceda2","en","Kandinsky 6.0 Open-Sourced: 29B Synced Audio-Video Models","Kandinsky 6.0 open-sourced: 3B and 29B models generate 5-second video with 44kHz synced audio and lip-sync; RL post-training cut WER 47%. MIT.","Video generation models have long shipped with an awkward default: silence. Voiceover, sound effects and ambient audio get bolted on later with TTS tools and editors, and lip-sync is manual work. On October 4, Kandinsky Lab released the paper \"Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation\" ([arXiv:2610.05608](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.05608)), and two days later open-sourced the weights, code and diffusers integration under the MIT license — turning synchronized audio-video generation from a closed-flagship feature into a public capability.\n\n## Two sizes, one pass\n\nThe release is a pair of models: Lite at 3B parameters and Pro at 29B. Both directly output 5-second clips with synchronized 44kHz audio and lip-sync, in text-to-audio-video (T2AV) and image-to-audio-video modes; a separate 1.4B super-resolution diffusion model upscales results to Full-HD (1920×1080). The paper lists more than 80 authors.\n\n## Dual-stream CrossDiT\n\nThe architectural story is the dual-stream design: a pretrained video stream building on Kandinsky 5.0's video capabilities, plus an audio stream trained from scratch — first on large-scale audio-only corpora, then jointly with the video stream on paired audio-video data. Bidirectional cross-attention aligns the two streams in time and semantics while preserving unimodal fidelity. Pretraining is followed by supervised fine-tuning, RL-based post-training and distillation. Per the official numbers, RL post-training cut Pro's speech word error rate (WER) by 47% — the hardest number in this release.\n\nOn the team's own VABench benchmark, Pro beats LTX 2.5 and the Lite sibling on speech quality, audio aesthetics, AV alignment, lip-sync and visual realism; in human evaluation Pro clearly outperforms Kandinsky 5.0 Pro and is preferred over LTX 2.5 on visual quality, motion realism and artifact reduction, with a statistically significant edge in speech quality. These are vendor-run evaluations — treat them as official claims, not independent ones.\n\n## Consumer GPUs and day-one ecosystem\n\nThe README's measured numbers are refreshingly honest: on the non-distilled Lite, a 5-second clip takes about 437 seconds on an RTX 4090 and 239 seconds on an H100; Pro at Full-HD takes about 402 seconds on an H100 and roughly 854 seconds on an RTX 5090 with block offload. Slow, but consumer cards genuinely run it. On release day (October 6) the ecosystem landed together: ComfyUI node packs, vLLM-omni support and a Hugging Face Space demo went live the same day; the GitHub repo passed 120 stars within a day, and 8 model repos on Hugging Face already reference the paper.\n\n## So what\n\nSynchronized audio-video generation has mostly been closed-flagship territory; open-source stacks assembled it from a video model plus TTS, with alignment and lip-sync as chronic pain points. Kandinsky 6.0 drops the full stack — two base models, super-resolution, distilled variants — under MIT at once, which is unusually friendly to local deployment and derivative work. The open question is length: 5 seconds is the current ceiling, and stretching synchronized generation to minutes is the next real fight.","kandinsky-6-0-video-audio-generation-open-source","2026-10-06T19:15:00Z","2026-10-06T19:10:43.708296Z","2026-10-06T19:10:43.708307Z",true,"agent",433,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"5612d186-46ee-4509-9a93-94045ba004ae","LTX-2.5 开放权重视频模型:4K 反而在 Fast 端点,EXR 色彩管线也焊进去了","ltx-2-5-open-weights-video","2026-08-18T15:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"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":75,"title":76,"news_slug":77,"published_at":78},"24d56562-4886-4169-8aa9-773f5fa34e2f","GRACE 把视频潜空间压 8 倍:提速 11 倍,跑分反超","grace-latent-compression-video-diffusion","2026-10-08T15:10:59+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"40d7bf26-810e-4286-900f-f343fdb28ad5","ID-Forcing:免训练把5秒视野外推到240秒","id-forcing-long-video-kv-provenance","2026-10-07T23:09:14+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"af21d26d-d9cd-4526-ac4a-66366a45848c","AV-GRPO:8张A800给22B音视频模型做RL后训练","av-grpo-audio-video-diffusion-rl","2026-09-27T17:08:13+00:00"]