arXiv:2503.10678cs.CV2025-03被引 2

用文本描述生成视频中指定物体的透明度图,实现精准交互抠图。

VRMDiff: Text-Guided Video Referring Matting Generation of Diffusion

  • 将抠图转为视频生成任务,利用扩散模型对齐文本与视频语义。
  • 提出新型损失函数,显著提升不同物体间的区分能力。
  • 首个包含文本、视频和实例级透明度图的大规模数据集,适合视频编辑研究者使用。

我们提出了一项新任务——视频指代抠图,即通过输入一段指代描述,获取视频中特定实例的透明度图(alpha matte)。我们将抠图这一密集预测任务视为视频生成问题,利用视频扩散模型的文本-视频对齐先验,生成时序连贯且与语义实例紧密相关的透明度图。此外,我们设计了一种新的隐空间构建损失(Latent-Constructive loss),进一步增强对不同实例的区分能力,实现更可控的交互式抠图。同时,我们构建了一个大规模视频指代抠图数据集,包含10,000个视频,是首个同时包含文本描述、视频序列和实例级透明度图的数据集。大量实验验证了方法的有效性。代码与数据集已公开于 https://github.com/Hansxsourse/VRMDiff。

原文摘要 · Abstract (English)

We propose a new task, video referring matting, which obtains the alpha matte of a specified instance by inputting a referring caption. We treat the dense prediction task of matting as video generation, leveraging the text-to-video alignment prior of video diffusion models to generate alpha mattes that are temporally coherent and closely related to the corresponding semantic instances. Moreover, we propose a new Latent-Constructive loss to further distinguish different instances, enabling more controllable interactive matting. Additionally, we introduce a large-scale video referring matting dataset with 10,000 videos. To the best of our knowledge, this is the first dataset that concurrently contains captions, videos, and instance-level alpha mattes. Extensive experiments demonstrate the effectiveness of our method. The dataset and code are available at https://github.com/Hansxsourse/VRMDiff.

视频生成扩散模型抠图文本引导

Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。