FreeDeshadow removes only the shadow you ask for: a dot, a line or a sentence picks the object, and non-shadow PSNR reaches 51.078 dB
Synopsis
Existing shadow removal either processes the whole image or depends on a precise shadow mask, so it cannot remove only the shadow cast by a user-specified object. FreeDeshadow shifts the interaction target from the shadow region to the casting object, accepting dot, line and object-mask visual prompts plus attribute, spatial-relation and state textual prompts; a generator produces a shadow-free proposal, and a local refiner establishes object-to-shadow correspondence and confines changes to the target shadow. On PCSRD+, non-shadow PSNR is 51.078 dB under visual prompts and 47.955 dB under textual prompts.
Interpretation
Shadow removal moves from whole-image or shadow-mask operation to operation on a user-specified object: only that object's cast shadow is removed, while the object, other objects' shadows and the background stay unchanged. Earlier shadow-aware methods need a precisely annotated shadow mask as input, and a mask says where shadows are but not which object casts them, so they cannot make object-level choices in multi-object scenes. On PCSRD+ with visual prompts, non-shadow PSNR is 51.078 dB versus 39.553 dB for the task-adapted Qwen-Image-Edit-2511 plus LoRA baseline; with textual prompts it is 47.955 dB versus 39.108 dB.
The system first interprets the prompt to understand the target object and generates an overall shadow-free candidate image, then pastes back only the part of that candidate belonging to the object's cast shadow. The generator builds on Qwen-Image-Edit, encoding image and prompt into a unified multimodal representation with a frozen Qwen2.5-VL encoder and fine-tuning the diffusion backbone with LoRA; the local refiner turns that representation into sparse shadow prompts and uses SAM2 multi-scale features to link the object to its spatially separated shadow. Evaluation runs on PCSRD+, which contains 154,571 quadruples at 1024×1024 resolution and covers dot, line and object-mask visual prompts plus attribute, spatial-relation and state textual prompts.
Perspective
The demonstrated scope is six prompt types on the PCSRD+ dataset, plus a cross-dataset test on WSRD without fine-tuning and several real-world images; the intended readers are image-editing and computational-photography researchers who need object-level shadow editing.
On the textual-prompt side, shadow-region PSNR is 21.553 dB, slightly below the 21.895 dB of Qwen-Image-Edit-2511 plus LoRA, with the advantage appearing in non-shadow and full-image metrics; cross-dataset generalization is shown only as qualitative results on WSRD.
