PulseAugur
EN
LIVE 07:32:31

Sparkle dataset and benchmark advance AI video background replacement

Researchers have introduced Sparkle, a new dataset and benchmark designed to improve instruction-guided video background replacement. The dataset contains approximately 140,000 video pairs across five themes, addressing the scarcity of high-quality training data that has hindered previous models like Kiwi-Edit. Sparkle utilizes a novel pipeline for decoupled foreground and background guidance synthesis, leading to more natural and temporally consistent scene changes. Experiments show that models trained on Sparkle outperform existing baselines on both the new Sparkle-Bench and the OpenVE-Bench. AI

IMPACT Enhances capabilities for video editing tools, potentially improving content creation in film and advertising.

RANK_REASON This is a research paper introducing a new dataset and benchmark for video background replacement.

Read on arXiv cs.CV →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

Sparkle dataset and benchmark advance AI video background replacement

COVERAGE [2]

  1. arXiv cs.CV TIER_1 English(EN) · Ziyun Zeng, Yiqi Lin, Guoqiang Liang, Mike Zheng Shou ·

    Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance

    arXiv:2605.06535v1 Announce Type: new Abstract: In recent years, open-source efforts like Senorita-2M have propelled video editing toward natural language instruction. However, current publicly available datasets predominantly focus on local editing or style transfer, which large…

  2. arXiv cs.CV TIER_1 English(EN) · Mike Zheng Shou ·

    Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance

    In recent years, open-source efforts like Senorita-2M have propelled video editing toward natural language instruction. However, current publicly available datasets predominantly focus on local editing or style transfer, which largely preserve the original scene structure and are…