Feyn公司推出了MultiMatte，一款可通过文本指令进行背景移除的AI模型1。用户可采用自然语言提示词指定需要保留的对象，模型随后会移除图像中的其他所有元素1。
该模型基于Meta的SAM 3概念可提示检测器构建1，采用alpha遮罩技术而非二进制掩码，能更好地处理毛发、毛皮及运动模糊等模糊边界区域1。在性能表现上，MultiMatte在DIS5K基准的S-measure指标从0.674提升至0.908，相对提升34.6%1；在DUT-OMRON基准上的表现也从0.792提升至0.901，相对提升13.7%1。该模型已在GitHub上开源发布，同时提供在线演示供用户体验1。
Feyn has released MultiMatte, an open-source image background removal model that leverages natural language prompts to specify which objects to preserve while removing all other elements from images 1. Built on the foundational concepts of Meta's SAM 3 architecture, the model employs alpha matting rather than binary masking, enabling superior handling of complex boundaries such as hair, fur, and motion blur 1.
The model demonstrates substantial performance improvements across standard benchmarks 1. On the DIS5K dataset, the S-measure metric improved from 0.674 to 0.908, representing a relative gain of 34.6%, while performance on the DUT-OMRON benchmark increased from 0.792 to 0.901, a relative improvement of 13.7% 1. MultiMatte represents Feyn's second-generation background removal offering and is available as open-source software on GitHub at https://github.com/feyninc/nobg, with an online demonstration accessible at https://usefeyn.com/multimatte/ 1.
评论
还没有评论，欢迎留下第一条。