Microsoft's LensErase: a foundation model for erasing anything from images
microsoft · hf · 2026-10-15
Microsoft releases LensErase, a general-purpose mask-based "erase anything" foundation model covering objects, visual text, watermarks, visual effects, multi-object removal, and varied scales/resolutions.
Key contributions:
- LensErase-100M: 101.07M paired source-mask-target examples across diverse erase scenarios
- Pretrained from the Lens text-to-image model, adapting its generative prior to mask-based erasing
- Mask-Aware RoPE: 3D rotary encoding that injects mask info into source-image tokens without a separate mask token sequence
- Refined via SFT + RL with contract-guided, region-aware rewards; TraceFusion at inference preserves fidelity outside the edit region
Achieves substantial gains on six public benchmarks and LensErase-Bench under both reference-based and VLM evaluation.
More from Multimodal
- ATAS framework uses world-state memory to keep long-form agentic video generation consistent — zju · 2026-10-15
- Building a 'hypnotic piano groove' skill for AI music systems, modeled on Todd Terje — doodlestein · 2026-10-15
- AI-Made Kart Racer Acorn Rally Adds Moon and Bayou Tracks on Spawn — AIandDesign · 2026-10-15
- MCP for Blender hits 30.9K stars in 18 months, letting Claude model and light 3D scenes — sidahuj · 2026-10-15
- Runway CEO reacts to claim it brought DeepMind's world model theory to life first — c_valenzuelab · 2026-10-15
- Zelda Stop-Motion Short Made with MiniMax H3, Plus a Trick to Fix Oily AI Skin — dramaton42 · 2026-10-15