DiffDecompose：基于扩散 Transformer 的 Alpha 合成图像逐层分解
DiffDecompose: Layer-Wise Decomposition of Alpha-Composited Images via Diffusion Transformers
把扩散 Transformer 当作后验采样器来解 alpha 合成图像的逐层分解，其 In-Context Decomposition 与 Layer Position Encoding Cloning 可直接迁移到去遮挡、透明物体分离、图层化编辑等需要多图层一致输出的视觉任务。
原文依据 DiffDecompose performs In-Context Decomposition, enabling the model to predict one or multiple layers without per-layer supervision
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩散模型近期在物体移除等众多生成任务上取得了巨大成功。然而，现有的图像分解方法受制于掩码先验依赖、静态物体假设以及数据集匮乏，难以解耦半透明或透明的图层遮挡。本文深入探索一项新任务：Alpha 合成图像的逐层分解，目标是在半透明/透明 alpha 图层非线性遮挡条件下，从单张重叠图像中恢复出各个组成图层。为应对图层歧义、泛化性与数据稀缺等挑战，我们首先提出 AlphaBlend，这是首个面向透明与半透明图层分解的大规模高质量数据集，包含六个特性各异的子任务（例如半透明眩光去除、半透明细胞分解、玻璃器皿分解）。在该数据集基础上，我们提出 DiffDecompose，一个基于扩散 Transformer 的框架，它以输入图像、语义提示与混合类型为条件，学习可能图层分解的后验分布。DiffDecompose 并不直接回归 alpha 蒙版，而是执行上下文分解（In-Context Decomposition），使模型无需逐层监督即可预测一个或多个图层，并引入图层位置编码克隆（Layer Position Encoding Cloning）以维持图层间的像素级对应关系。在所提出的 AlphaBlend 数据集与公开的 LOGO 数据集上的大量实验验证了 DiffDecompose 的有效性。代码将公开于 https://github.com/Wangzt1121/DiffDecompose。
英文原文摘要 1629 字符
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, containing six subtasks with different characteristics (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. Code will be publicly available at https://github.com/Wangzt1121/DiffDecompose.