DiffDecompose: Layer-Wise Decomposition of Alpha-Composited Images via Diffusion Transformers

TL;DR

This paper introduces DiffDecompose, a diffusion Transformer framework for layer-wise decomposition of alpha-composited images, supported by a new large-scale dataset AlphaBlend, enabling better handling of semi-transparent occlusions.

Contribution

The paper presents a novel diffusion Transformer-based approach and a large-scale dataset for transparent and semi-transparent layer decomposition from single images.

Findings

DiffDecompose outperforms existing methods on AlphaBlend and LOGO datasets.

The AlphaBlend dataset supports six real-world decomposition subtasks.

DiffDecompose enables in-context layer prediction without per-layer supervision.

Abstract

Diffusion models have recently motivated great success in many generation tasks like object removal. Nevertheless, existing image decomposition methods struggle to disentangle semi-transparent or transparent layer occlusions due to mask prior dependencies, static object assumptions, and the lack of datasets. In this paper, we delve into a novel task: Layer-Wise Decomposition of Alpha-Composited Images, aiming to recover constituent layers from single overlapped images under the condition of semi-transparent/transparent alpha layer non-linear occlusion. To address challenges in layer ambiguity, generalization, and data scarcity, we first introduce AlphaBlend, the first large-scale and high-quality dataset for transparent and semi-transparent layer decomposition, supporting six real-world subtasks (e.g., translucent flare removal, semi-transparent cell decomposition, glassware decomposition). Building on this dataset, we present DiffDecompose, a diffusion Transformer-based framework that learns the posterior over possible layer decompositions conditioned on the input image, semantic prompts, and blending type. Rather than regressing alpha mattes directly, DiffDecompose performs In-Context Decomposition, enabling the model to predict one or multiple layers without per-layer supervision, and introduces Layer Position Encoding Cloning to maintain pixel-level correspondence across layers. Extensive experiments on the proposed AlphaBlend dataset and public LOGO dataset verify the effectiveness of DiffDecompose. The code and dataset will be available upon paper acceptance. Our code will be available at: https://github.com/Wangzt1121/DiffDecompose.