O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

Lightricks Open-Sources LTX-2.5: Generate 10-Second Videos in 6.8 Seconds with Native Multi-Shot and Audio

LTX, the open-world model company spun out of Lightricks, has released LTX-2.5, an open-weights video model that generates a 10-second 720p clip in 6.8 seconds on NVIDIA GB200 hardware. It supports text, image, and video inputs, produces synchronized audio and video, and offers native multi-shot generation. Free for organizations under $10M ARR.

Publicado em 13/08/2026, 04:16:45Última atualização 13/08/2026, 04:16:45Publicação após revisão automática

LTX, the open-world model company spun out of Lightricks, has officially released LTX-2.5, a new open-weights video and world model. The model is now available on Hugging Face and comes natively integrated into ComfyUI through a strategic partnership, with hosted API access also offered.\n\nPerformance-wise, LTX-2.5 delivers remarkable speed. In internal tests on two NVIDIA GB200 superchips, the model generates a 10-second 720p video in just 6.8 seconds—faster than real-time playback. It accepts text, image, and video inputs, and generates synchronized audio and video within a single model, eliminating the need for separate audio pipelines. Its native multi-shot generation renders an entire sequence as one output, preserving character, scene, and voice consistency across cuts rather than stitching together independently generated shots.\n\nOn the licensing side, LTX continues its open strategy: organizations with annual recurring revenue (ARR) below $10 million can use the model for free; larger companies must negotiate a commercial license. LTX views this 'free first, pay later' model as a way to attract developers and convert them to paying customers as they scale. For hosted generation, LTX-2.5 Fast costs $0.09 per second for 720p video with audio, or about $0.90 for a 10-second clip.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

LTX发布LTX-2.5模型：原生集成ComfyUI，主打开源与高效推理_ZAKER新闻app.myzaker.com · Complementar

速度与质量表现：数据背后的硬件前提 LTX 公布的性能数据显示，LTX-2.5 在两台 NVIDIA 顶级 GB200 芯片上自托管测试时，生成一段 10 秒、720p 分辨率的视频仅需 6.8 秒，速度快于实时播放。若通过 LTX 托管 API 以 1080p 分辨率完成相同任务，耗时为 23.7 秒。值得注意的是，上述极速成绩依赖于远超常规部署水平的顶级硬件。在横向对比中，LTX 基于端到端测量数据称，Google Gemini Omni Flash 耗时 52 秒，xAI Grok 1.5 耗时 63 秒，而字节跳动 Seedance 2.5 和快手 Kling 3.0 Pro 的耗时分别高达 317 秒和 398 秒。在质量盲测中，LTX-2.5 以 67% 的胜率领先，高于 Seedance 2.5（65%）和 Gemini Omni Flash（55%）。但 LTX 强调，这些数据均由供应商报告或委托测量，未经独立验证，仅作为方向性参考，企业应结合实际工作负载进行测试。战略抉择：为何押注 " 开源权重 " 而非封闭 API LTX 联合创始人兼 CEO Zeev Farbman 指出，视频和世界模型的应用场景远比语言模型广泛，封闭 API 无法满足企业构建定制化工作流的需求。" 开放性并非慈善，而是我们的战略。"Farbman 表示，通过提供免费开源权重吸引开发者，待客户业务规模化（ARR 超过 1000 万美元）后再转化为许可收入，是更符合行业特性的商业模式。 [...] 星途科讯 19小时前 Lightricks 分拆出的开放世界模型公司 LTX 今日正式发布 LTX-2.5。这是一款开源权重的视频及 " 世界 " 模型，通过与 ComfyUI 的战略合作，实现了在该节点式工作流工具中的原生集成。目前，该模型权重已在 Hugging Face 上线，并内置于 ComfyUI 中；用户也可通过 LTX API 调用托管服务。在商业化策略上，LTX 采取分层许可模式：年经常性收入（ARR）低于 1000 万美元的组织可免费使用，大型公司则需协商许可协议。LTX 表示，其模型累计下载量已突破 3300 万次，成为市场上应用最广泛的 " 开放世界 " 模型系列。重构生成管道，聚焦效率与多模态能力 LTX-2.5 并非在旧核心上简单