Our approach involves steps:
1️⃣ Sliding window inpainting (Qwen Controlnet).
2️⃣ Object Verification (Grounded-SAM-2).
3️⃣ Human Preference ranking (ImageReward).
Repeat for 30k scenes (Places365) × 50 object categories (COCO) → 30M annotations, fully automated 🤖
Results: HiddenObjects.