BFS: BACK-TO-FRONT LAYERED IMAGE SYNTHESIS VIA KNOWLEDGE TRANSFER

Kyoungkook Kang, Gyujin Sim, Sunghyun Cho

Keywords: Layer synthesis, Generative Model

TL;DR: We tackle layered image synthesis with a background-to-foreground approach and overcome data scarcity by leveraging knowledge distillation from composite synthesis.

Abstract:

Layered images have long served as a crucial representation for creative editing, and the advent of large-scale generative models has recently spurred interest in their automatic generation. Nevertheless, existing approaches remain limited. Decomposition-based methods often struggle to achieve clean separation of layers in complex scenes, while generation-based methods face challenges stemming from their training data construction pipeline, which lead to suboptimal visual quality and limited scene diversity. In this paper, we propose BFS, a novel generation-based framework for layered image synthesis. We approaches the task through the highly practical background-to-foreground formulation. Specifically, given a background layer and user guidance, it synthesizes a foreground layer that incorporates visual effects such as shadow and reflection while harmonizing with the background to form a coherent composite image. Since constructing suit-
able training data is difficult, we instead leverage the comparatively easy-to-learn knowledge of composite synthesis for the foreground synthesis. To this end, we design a dual-branch framework that jointly generates a composite image and a foreground layer, enabling knowledge transfer through bidirectional information exchange between the two branches. To promote this transfer, we also propose a two-stage training scheme that does not rely on ground-truth foreground layer, main dataset bottleneck. Extensive experiments show that BFS produces high-quality layered images, consistently outperforming prior methods.

Primary Area: generative models

Submission Number: 6843