Core Insights - The arrival of the rainy season is boosting the production and sales of edible mushrooms in Yunnan, with a significant increase in transaction volume at the Yunnan Mushrooms Trading Center in Kunming [2] - The number of registered edible mushroom-related enterprises has been steadily increasing, with 64,000 new registrations this year and a total of 720,400 existing enterprises in China [3] Group 1: Current Market Situation - The transaction volume of popular mushrooms such as porcini, chicken of the woods, matsutake, and green-headed mushrooms has surged, leading to a decrease in prices and attracting many local residents and tourists [2] - The edible mushroom industry has created new job opportunities, including roles such as mushroom washers, scrapers, and identifiers [2] Group 2: Enterprise Registration Trends - There are currently 720,400 edible mushroom-related enterprises in China, with a peak registration of 197,900 enterprises in 2023, and an expected 157,300 registrations in 2024, indicating a high level of activity in the sector [3] - Enterprises established within the last 1-3 years make up the largest segment, accounting for 43.27% of the total, while those established within the last year and over 10 years represent 18.89% and 4.92%, respectively [4] Group 3: Industry Classification - Over 80% of edible mushroom-related enterprises fall under the categories of agriculture, forestry, animal husbandry, and fishery, with this segment accounting for 80.48% of the total [5] - Other significant classifications include wholesale and retail (7.35%) and scientific research and technical services (6.79%) [5]

机器之心发布 机器之心编辑部 视频是信息密度最高、情感表达最丰富的媒介之一，高度还原现实的复杂性与细节。正因如此，视频也是编辑难度最高的一类数字内容。在传统的视频编辑流程 中，若要调整或替换主体、场景、色彩或是移除一个物体，往往意味着无数帧的手动标注、遮罩绘制和精细调色。即使是经验丰富的后期团队，也很难在复杂场 景中保持编辑内容的时间一致性。 近年来，生成式 AI 尤其是扩散模型与多模态大模型的快速迭代，为视频编辑带来了全新的解题思路。从早期基于规则的特效工具，到目标识别与自动分割，再到 基于文本指令的视频生成与重绘，尽管 AI 已经为视频编辑带来了效率与可控性的双重提升，但在精度要求较高的场景中仍存在一系列挑战，例如当前很多零样本 方法在处理连续视频帧时容易造成画面闪烁；对于背景复杂或多目标场景，可能会出现错位、模糊或语义偏差。 针对于此，北京大学相机智能实验室（施柏鑫团队）联合 OpenBayes贝式计算，以及北京邮电大学人工智能学院模式识别实验室李思副教授团队，共同提出了一种 结合草图与文本引导的视频实例重绘方法 VIRES，支持对视频主体的重绘、替换、生成与移除等多种编辑操作。该方法利用文本生成视频模 ...