Core Concepts
Decoupling generative object compositing into identity preservation and background alignment stages significantly improves realism and fidelity.
Abstract
IMPRINT introduces a novel two-stage framework for generative object compositing. The first stage focuses on context-agnostic identity preservation, while the second stage harmonizes the object with the background. Extensive experiments show IMPRINT outperforms existing methods in identity preservation and composition quality. The model incorporates shape-guidance for user control over compositing.
Stats
IMPRINT significantly outperforms existing methods in identity preservation and composition quality.
The first stage is trained on 1,409,545 pairs and validated on 11,175 pairs from MVImgNet.
The second stage is fine-tuned on a mixture of image datasets and video datasets, including a training set of 217,451 pairs.