One-Step Diffusion Model OSOR Cuts Object Removal Times by Up to 30x

Removing objects from images convincingly has long meant waiting through dozens of diffusion sampling steps to get a clean, artifact-free result. A newly published research paper describes a model called OSOR, short for One-Step Object Removal, that claims to collapse that process into a single diffusion step, running between 4x and 30x faster than multi-step baselines while matching or exceeding their visual quality.

The paper, posted to arXiv on June 26, 2026, tackles a problem that has dogged fast diffusion models: single-step training is notoriously unstable, and models that skip the iterative refinement process tend to produce blurry edges or leftover ghosting where an object used to be. OSOR's authors address this with what they call an occupancy-guided discriminator, a component designed to give the model precise feedback on object boundaries during training so that single-step outputs stay sharp and coherent rather than degrading into visual noise.

A second piece of the architecture, described as an alpha head, draws on pretrained diffusion models to predict which regions of an image actually need to be removed. This matters in practice because real-world masks marking the object to erase are frequently imprecise, either too tight or bleeding into surrounding pixels. By predicting removal regions rather than strictly trusting the input mask, OSOR is designed to compensate for that imperfection instead of propagating it into the final image.

Building Cleaner Training Data at Scale

Fast, stable inference is only half the story. OSOR's training also depends on a semantic-anchored verification pipeline, referred to as SAVP, which filters out noisy or mismatched instruction-based image triplets before they reach the model. The authors used SAVP to build CORNE, a curated dataset of 280,000 verified removal pairs, which the paper positions as a cleaner and more reliable foundation than typical scraped or automatically generated training triplets.

To test how OSOR handles harder edge cases, the team also assembled two new evaluation sets: AnimeEraseBench, focused on removing objects from stylized illustration and anime-style imagery, and TextEraseBench, aimed at scenarios where text needs to be cleanly erased from a scene. Both benchmarks target categories of images where object boundaries and semantic context are historically difficult for generative inpainting and removal systems to get right.

The authors state that code and supporting resources for OSOR are being made available, which would let other researchers verify the speed and quality claims independently and potentially plug the occupancy-guided discriminator or alpha head into other diffusion pipelines.

If the reported gains hold up under independent testing, the practical implications are straightforward: object removal tools used in photo editing software, e-commerce image cleanup, and video post-production could shift from a multi-second or multi-step operation to something closer to real-time, without users having to trade quality for speed. That kind of speedup matters most in production settings where removal has to run at scale across large volumes of images rather than one photo at a time.