Adobe Research's Object-WIPER Removes Objects and Their Visual Traces from Video Without Retraining

Object-WIPER, a training-free research method from Adobe Research and UT Dallas accepted to CVPR 2026, goes beyond simple mask-and-fill object removal by also erasing associated visual effects — shadows, reflections, and translucent traces — that typically reveal an edit. The system leverages a pre-trained text-to-video diffusion transformer (DiT), steering it with a user-supplied object mask and a short text description (e.g., "a dog and its shadow") to locate and erase both the object and its scene-wide traces via visual-text cross-attention, without any task-specific fine-tuning.

To reconstruct clean footage, the method inverts the video through the DiT to extract structured noise, replaces masked tokens with fresh Gaussian noise, and reintroduces preserved background tokens during denoising to maintain temporal consistency across frames. The team also introduces WIPER-Bench, a real-world benchmark and accompanying metric covering common effect types such as shadows and reflections, designed to evaluate scene coherence after removal. Adobe has noted this is academic research and not a shipping product feature.