arXiv

AlbedoEdit: Unified Instance-Level Video Editing with Albedo Guidance

Title: AlbedoEdit: Unified Instance-Level Video Editing with Albedo Guidance

Abstract:

While video generative models have made significant strides in producing photorealistic video sequences, unlocking their full potential for creative downstream applications demands more sophisticated capabilities. Specifically, fine-grained instance-level video editing—encompassing tasks such as object insertion and removal, as well as texture modification—has become a critical yet difficult challenge. Current methods tend to fall short; they either rely on unified generative frameworks that offer only coarse semantic control or employ specialized, task-specific models that restrict flexibility and applicability in varied real-world contexts.

To overcome these constraints, we introduce AlbedoEdit, a comprehensive generative video editing framework designed to seamlessly handle object insertion, removal, and texture editing within a single architecture. The core innovation behind AlbedoEdit lies in the utilization of intrinsic albedo maps. Because albedo is invariant to lighting conditions and excludes effects such as specularity, shadows, and inter-reflections, it serves as an intuitive and effective tool for users to define precise appearance adjustments.

Built on top of video foundation models, AlbedoEdit is fine-tuned to transform source RGB videos into edited versions, guided by a user-modified albedo map of the first frame. The model was trained on a novel paired synthetic dataset that includes examples of all three editing tasks. Through this training, AlbedoEdit implicitly acquires the ability to blend edited content harmoniously and replicate complex visual phenomena resulting from these edits, such as specular highlights, soft shadows, and mirror reflections. Our evaluations demonstrate that AlbedoEdit outperforms existing state-of-the-art video editing techniques, achieving superior results in both qualitative assessments and quantitative metrics. For more details, please visit the project webpage: https://vcai.mpi-inf.mpg.de/projects/AlbedoEdit/.


Source: arXiv Generated at: 2026-06-02 00:00:00 UTC

Related Articles

Law’s Billable Hour Is Being Shredded by AI
Bloomberg

Law’s Billable Hour Is Being Shredded by AI

AI is dismantling the billable hour by automating routine legal tasks. This technological shift threatens the traditiona...

Iran War: Trump Tries to Stop Israel’s Lebanon Push | The Opening Trade 6/2/2026
Bloomberg

Iran War: Trump Tries to Stop Israel’s Lebanon Push | The Opening Trade 6/2/2026

SoftBank in Early Talks to Back $800 Million Agile Robots Round
Bloomberg

SoftBank in Early Talks to Back $800 Million Agile Robots Round

SoftBank is in early talks to back Agile Robots’ $800 million funding round. The Japanese tech giant is currently in pre...

Amundi Is Diversifying Risk Via Commodity Currencies, Gold
Bloomberg

Amundi Is Diversifying Risk Via Commodity Currencies, Gold

Amundi diversifies risk by investing in commodity-linked currencies and gold. This strategy hedges against market volati...

Reuters

Marvell Technology surges after Nvidia's Huang calls it 'next trillion-dollar company'

Marvell Technology shares surged after Nvidia CEO Jensen Huang labeled the firm the “next trillion-dollar company.”

Russia Says It Found Foreign Spyware on Top Officials’ Phones
Bloomberg

Russia Says It Found Foreign Spyware on Top Officials’ Phones

Russia’s FSB claims to have discovered foreign spyware on senior officials’ phones. Moscow attributes the intrusion to h...