FreeDeshadow: Object-Aware Controllable Shadow Removal via Multimodal Prompts

Authors: Zhiliang Wu, Jihao Gu, Kun Li, Hanyu Xuan, Hehe Fan, Meng Wang, Weisi Lin

Subjects: Computer Vision and Pattern Recognition (cs.CV)

arXiv: 2611.14608 [cs.CV] — Submitted 10 Nov 2026

Project Page: https://wzlwlm13.github.io/FreeDeshadow

Overview

Conventional shadow removal approaches fall into two broad categories: methods that perform global shadow removal across an entire image, and methods that depend on precise shadow masks to guide restoration. In practical editing scenarios, however, users often want to eliminate only the shadow cast by one specific object while leaving the rest of the scene untouched. Neither paradigm supports this level of selective control.

To address this gap, the authors introduce object-aware controllable shadow removal, a new task in which a user identifies a target object through a simple prompt and the system removes only that object's cast shadow. The remaining content of the image is preserved. The work is released as FreeDeshadow, a unified framework that supports both visual and textual prompts.

Key Challenges

Selective shadow removal is difficult because of an inherent object–shadow asymmetry: an object and the shadow it casts are spatially separated in the image. This separation creates two core problems:

Correspondence. The system must accurately associate the user-specified object with its spatially distant cast shadow.
Localization. Restoration must be confined to the target shadow region so that unrelated areas remain unchanged.

Method

FreeDeshadow is built around two cooperating components:

Object-aware generator. Interprets the user prompt and produces a high-quality shadow-free proposal, establishing the overall edit.
Shadow-aware local refiner. Establishes the correspondence between the object and its cast shadow and constrains generated changes to the target shadow region.

This design allows the framework to handle both visual prompts (a dot, a line, or an object mask) and textual prompts that describe the target object by its attributes, spatial relations, or states.

Dataset: PCSRD+

To support training and evaluation, the authors construct PCSRD+, a large-scale dataset comprising 154,571 quadruples at 1024×1024 resolution. It covers six types of user prompts:

Visual: a dot, a line, and an object mask
Textual: descriptions of the target object based on attributes, spatial relations, or states

Results

Extensive experiments show that FreeDeshadow accurately removes the shadow of the target object under both visual and textual prompts while effectively preserving non-target regions. The method outperforms existing approaches in both shadow removal performance and controllability.

Citation

@misc{wu2026freedeshadow,
  title  = {FreeDeshadow: Shadow Removal for User-Specified Objects With Multimodal Prompts},
  author = {Wu, Zhiliang and Gu, Jihao and Li, Kun and Xuan, Hanyu and Fan, Hehe and Wang, Meng and Lin, Weisi},
  year   = {2026},
  eprint = {2611.14608},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

via ArXiv CV

FreeDeshadow: Object-Aware Controllable Shadow Removal via