[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ltx-2-5-open-weights-video":3,"news-related-5612d186-46ee-4509-9a93-94045ba004ae":38},{"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},"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","Lightricks 发布开放权重音视频生成模型 LTX-2.5:Fast 端点支持 4K\u002F50fps、最长 20 秒片段并默认输出同步音频,Pro 端点主打高保真;新增 Diffusion Fidelity Rendering、扩散视频解码器、原生多镜头与 EXR 专业色彩工作流,并附带可微调的 raw 预训练 checkpoint。许可证为 LTX-2.x Community License,年营收达 1000 万美元以上的实体需先获取付费商业许可。","# LTX-2.5 开放权重视频模型:4K 反而在 Fast 端点,EXR 色彩管线也焊了进去\n\n想用 LTX-2.5 生成 4K 视频?你得选那个标着「Fast」的端点,而不是「Pro」。这不是 bug:Lightricks 官方文档写得很清楚,Fast 端点用一部分保真度换取速度、4K 输出和最长 20 秒的片段,而质量优化的 Pro 端点只跑到 1080p、10 秒以内。这个反直觉的设计,恰好是理解 LTX-2.5 产品哲学的入口:它不是在卷「更大更强」,而是在把一个开放权重的视频模型,往真正的生产管线里塞。\n\n## 六个端点,一次生成,声音同步\n\nLTX-2.5 是 Lightricks 的开放权重音视频生成模型,以 LTX-2.x Community License 发布,在 fal 平台提供六个端点:文生视频、图生视频、音频生视频,各有质量优化的 Pro 与速度优化的 Fast 两个变体。所有变体默认输出同步音频,支持 16:9 和 9:16 画幅;图生视频还接受一个「结束帧」,让用户钉住镜头的落点。\n\n规格上,Fast 变体覆盖 720p\u002F1080p\u002F1440p\u002F2160p(4K),帧率 24\u002F25\u002F48\u002F50fps,片段最长 20 秒;Pro 变体是 720p\u002F1080p,24\u002F25\u002F50fps,片段 6\u002F8\u002F10 秒。参数背后的分工耐人寻味:Fast 负责快速迭代与 4K 交付,Pro 负责最终高保真输出。\n\n## 画质提升的账本:算力花在哪很重要\n\nLTX-2.5 引入 Diffusion Fidelity Rendering:把更多算力投给复杂场景,而不是平均撒到每一帧上——人群、快速运动、高密度细节在这些地方撑得住,而不是糊掉。这一特性运行在 Pro 端点上。\n\n官方 FAQ 认为,相对 2.3 版本保真度跃升「最大的单一贡献项」是 Diffusion Video Decoder:用扩散解码器替换普通 VAE 解码,人脸更锐、屏幕文字可读、快速运动拖影更少。\n\n叙事能力上的增量是原生多镜头(native multishot):一次生成产出多个连续镜头,角色、环境、光照、声音、风格在每次剪辑间保持一致。支撑它的是一个定制的 Gemma 4 12B 文本编码器,能在一个复杂提示词里追踪多个主体、动作、光照提示与镜头调度;Auto Duration 则在扩散开始前先「读懂」描述的动作,预测合适的片段长度。\n\n## 更「工业」的一步:EXR 与 raw checkpoint\n\n两个细节说明 Lightricks 盯上的是专业制片流程。其一,LTX-2.5 增加了原生 EXR 工作流,直接在 ACES、DaVinci Wide Gamut 等专业色彩空间内读写电影级 EXR,生成式编辑返回的也是 EXR——省掉了色彩管线中间那次有损的 8-bit 往返。\n\n其二,是随模型一同发布的 raw pretrained checkpoint:一个未经监督微调(SFT)的基底,官方给它的定位是向新数据与新目标做激进适配——机器人、合成音视频、工业数字孪生、私有领域模型。配合 LoRA 微调(fal 平台已托管 80 多个 LTX 微调版本),「开放权重」在这里不只是「能下载」,而是「能改造」。\n\n## 价格与许可证的边界\n\nfal 平台按秒计费:Fast 变体 720p 每秒 0.09 美元、1080p 每秒 0.13 美元、1440p 每秒 0.19 美元、4K 每秒 0.30 美元,各分辨率都包含原生音频;Pro 图生视频 720p\u002F1080p 分别为每秒 0.12\u002F0.17 美元。\n\n许可证方面需要泼一盆冷水:LTX-2.x Community License 不是 OSI 意义上的开源许可。它允许商业使用,但年营收达到 1000 万美元以上的实体需要先向 Lightricks 获取付费商业许可,且许可条款带有使用限制——例如不得用它训练竞争模型。\n\n## 所以呢\n\n闭源 API 视频模型拼的是「一句话出大片」的惊艳感,而 LTX-2.5 在拼另一件事:可自托管、可微调、能接进 ACES 色彩管线、能跑在自有基础设施上的**资产**。4K 放在 Fast 端点这个「反直觉」,本质是把高分辨率当成交付格式,把 Pro 留给画质本身——这是制片逻辑,不是 demo 逻辑。开放权重视频模型的下一轮竞争,可能不再由参数表决定,而由谁先被缝进工作流决定。\n\n原文与模型详情见 [fal.ai 的 LTX-2.5 页面](https:\u002F\u002Ffal.ai\u002Fltx-2.5)。","https:\u002F\u002Ffal.ai\u002Fltx-2.5","234f16d5-2704-4fbf-b70a-059c5164830d",[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},"d9e717f9-4bb8-4ac0-89f0-5d5b768a7839","en","LTX-2.5 open weights: 4K on the fast lane, EXR built in","Lightricks has released LTX-2.5, an open-weights audio-video generation model: Fast endpoints support 4K at up to 50fps with clips up to 20 seconds and synchronized audio by default, while Pro endpoints focus on fidelity. New features include Diffusion Fidelity Rendering, a Diffusion Video Decoder, native multishot, and an EXR professional color workflow, plus a raw pretrained checkpoint for fine-tuning. It ships under the LTX-2.x Community License, which requires a paid commercial license for entities with annual revenue of 10 million US dollars or more.","# LTX-2.5 Open-Weights Video Model: 4K Lives on the Fast Endpoints, and an EXR Color Pipeline Is Built In\n\nWant 4K output from LTX-2.5? You need the endpoint labeled \"Fast,\" not \"Pro.\" That is not a bug: Lightricks documentation states plainly that the Fast endpoints trade some fidelity for speed, 4K output, and clips up to 20 seconds, while the quality-optimized Pro endpoints top out at 1080p and 10 seconds. This counterintuitive design is a good entry point for understanding the product philosophy of LTX-2.5: it is not competing on \"bigger and stronger,\" but on pushing an open-weights video model into real production pipelines.\n\n## Six Endpoints, One Generation, Audio Included\n\nLTX-2.5 is Lightricks open-weights audio-video model, released under the LTX-2.x Community License and served on the fal platform through six endpoints: text-to-video, image-to-video, and audio-to-video, each in a quality-optimized Pro variant and a speed-optimized Fast variant. Every variant generates synchronized audio by default and supports 16:9 and 9:16 framing. Image-to-video also accepts an end frame, letting users pin where a shot finishes.\n\nOn specs, Fast covers 720p\u002F1080p\u002F1440p\u002F2160p (4K) at 24\u002F25\u002F48\u002F50fps with clips up to 20 seconds; Pro runs 720p\u002F1080p at 24\u002F25\u002F50fps with 6-, 8-, or 10-second clips. The division of labor is telling: Fast handles rapid iteration and 4K delivery, Pro handles final high-fidelity output.\n\n## The Fidelity Ledger: Where the Compute Goes Matters\n\nLTX-2.5 introduces Diffusion Fidelity Rendering, which puts more compute into complex scenes instead of spreading it evenly across every frame — crowds, fast motion, and dense detail hold together where they would otherwise soften. The feature runs on the Pro endpoints.\n\nAccording to the official FAQ, the single biggest contributor to the fidelity jump over version 2.3 is the Diffusion Video Decoder: it replaces plain VAE decoding with a diffusion-based one, so faces stay sharp, on-screen text stays legible, and fast motion shows fewer smears.\n\nThe narrative-side increment is native multishot: a single generation yields multiple connected shots that hold character, environment, lighting, voice, and style across every cut. Supporting it is a custom Gemma 4 12B text encoder that tracks multiple subjects, actions, lighting cues, and camera direction through a complex prompt, while Auto Duration reads the described action to pick the right clip length before diffusion begins.\n\n## A More Industrial Step: EXR and the Raw Checkpoint\n\nTwo details show Lightricks is aiming at professional production. First, LTX-2.5 adds a native EXR workflow that reads and writes cinema-grade EXR inside professional color spaces including ACES and DaVinci Wide Gamut, with generative edits returning EXR — no lossy 8-bit round-trip in the middle of a color pipeline.\n\nSecond, the model ships with a raw pretrained checkpoint: a non-SFT base built for aggressive adaptation toward new data and objectives — robotics, synthetic AV, industrial digital twins, private domain models. Combined with LoRA training (fal already hosts more than 80 LTX fine-tunes), \"open weights\" here means not just downloadable, but modifiable.\n\n## Price and License Boundaries\n\nPricing on fal is per second: the Fast variant costs 0.09 USD per second at 720p, 0.13 USD at 1080p, 0.19 USD at 1440p, and 0.30 USD at 4K, with native audio included at every resolution; Pro image-to-video is 0.12 USD per second at 720p and 0.17 USD at 1080p.\n\nOn licensing, a bucket of cold water: the LTX-2.x Community License is not an OSI open-source license. It permits commercial use, but entities with annual revenue of at least 10 million US dollars must obtain a paid commercial license from Lightricks first, and the license carries use restrictions — such as not training competing models.\n\n## So What\n\nClosed-API video models compete on the wow factor of \"one prompt, one blockbuster.\" LTX-2.5 is competing on something else: being an asset that is self-hostable, fine-tunable, pluggable into an ACES color pipeline, and runnable on your own infrastructure. Putting 4K on the Fast endpoints is, at heart, treating high resolution as a delivery format while reserving Pro for image quality itself — that is production logic, not demo logic. The next round of competition among open-weights video models may be decided not by spec sheets, but by who gets stitched into a workflow first.\n\nSource and model details: [the LTX-2.5 page on fal.ai](https:\u002F\u002Ffal.ai\u002Fltx-2.5).","ltx-2-5-open-weights-video","2026-08-18T15:20:00Z","2026-08-18T15:06:58.756191Z","2026-08-18T15:06:58.756199Z",true,"agent",74,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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":46,"title":47,"news_slug":48,"published_at":49},"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":51,"title":52,"news_slug":53,"published_at":54},"18d2aa73-7244-4b10-b611-46475e17327e","ForgeWM开源:一步去噪72FPS的可玩世界模型,8张卡复现全流程","forgewm-few-step-playable-world-model","2026-08-24T21:10:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"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":61,"title":62,"news_slug":63,"published_at":64},"6e3002da-c1fd-4a6d-b903-4f65b976dd04","MiniMax H3 首个商用落点：美图 RoboNeo 接入背后,通用多模态模型的\"可编辑性\"才刚开始被检验","roboneo-minimax-h3-multimodal-editing","2026-08-03T18:02:02+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"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"]