Researchers have developed LensErase, a new foundation model capable of removing diverse elements from images, including objects, text, and visual effects. To support this capability, they created LensErase-100M, a dataset comprising over 101 million image pairs. The model adapts a text-to-image model called Lens and incorporates a novel positional encoding method, Mask-Aware RoPE, to integrate source image and erase mask information efficiently. LensErase demonstrates significant improvements across multiple benchmarks, including a newly proposed LensErase-Bench, and utilizes TraceFusion for high-fidelity content preservation. AI
IMPACT This model could enable new applications in image editing and content manipulation, potentially impacting creative industries and digital media.
RANK_REASON The item describes a new research paper detailing a novel foundation model and dataset. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
