StableDrag: Stable Dragging for Point-based Image Editing
Основні поняття
Designing a stable and precise drag-based editing framework, StableDrag, to enhance image manipulation through discriminative point tracking and confidence-based motion supervision.
Анотація
"StableDrag" is introduced as a framework for stable and precise point-based image editing. It addresses the limitations of existing dragging schemes by focusing on accurate point tracking and complete motion supervision. The proposed method aims to boost stability in long-range manipulation and ensure high-quality editing outcomes. By combining discriminative point tracking with confidence-based latent enhancement, two models, StableDrag-GAN and StableDrag-Diff, are instantiated for stable dragging performance. Extensive qualitative experiments and quantitative assessments on DragBench showcase the effectiveness of StableDrag in achieving stable and precise drag performance.
StableDrag
Статистика
DragDiffusion explores adapting the dragging scheme to diffusion models.
Discriminative point tracking model size: 1 × C × 1 × 1.
Learning rate for Adam optimizer: 0.01 for StableDrag-Diff, 0.001 for StableDrag-GAN.
Цитати
"Thanks to these unique designs, we instantiate two types of image editing models including StableDrag-GAN and StableDrag-Diff."
"Our contributions are summarized as follows: We propose a discriminative point tracking method."
"In detail, we utilize the tracking confidence score of the handle points to assess the quality of the current manipulation process."