A | IT之家 7 月 21 日消息,科技媒体 The Decoder 昨日(7 月 20 日)发布博文,报道称谷歌 DeepMind 发布 GenCeption 模型,将预训练的视频生成器重新用于深度估计和分割等经典计算机视觉任务。 (ECNS) -- China’s annual hydrogen production capacity exceeded 51 million metric tons by the end of 2025, up about 1.4% from a year earlier, while actual output rose 7.3% to more than 39 million metric tons, according to a report released Tuesday by the National Energy Administration. Annual production capacity for renewable hydrogen already in operation surpassed 250,000 metric tons, more than double the level recorded in 2024. The report said hydrogen applications were continuing to expand across the industrial, transport and energy sectors. China had about 32,000 fuel-cell vehicles on the road by the end of 2025, with larger-scale deployment being explored along freight corridors and at ports. Globally, 66 countries and regions had released hydrogen development strategies by the end of 2025, together accounting for more than 90% of global GDP. (By Helen Mo, intern Yang Hongran) 。
IT之家援引博文介绍,大语言模型在学习预测下一个 Token 的时候,在训练过程中往往需要吸收语法、世界知识和上下文关系等内容。
但是在计算机视觉领域,视觉模型缺少等效的训练方法,主要由专业模型主导,包括用于分割的“Segment Anything”和用于深度估计的“Depth Anything”,每个模型都使用其特定的架构。
B |
谷歌 DeepMind 团队为此提出 GenCeption 模型方案,尝试将一个“生成视频”的 AI 模型逆向改造成一个能“理解世界”的视觉分析引擎。