OverLay++: Dense-Overlap Layout-to-Image Generation Dataset

Key point Image models can place objects from boxes, but struggle with many overlapping and interacting objects.

However, existing methods still struggle with complex scenes containing many overlapping and interacting objects.

Key point Training data is a particular bottleneck: existing datasets lack dense, complex object interactions.

We argue that training data is a particular bottleneck: existing datasets lack examples with dense, complex object interactions.

Key point Older collections average only 1.78-3.93 objects per image, limiting supervision for dense scenes.

Existing datasets such as LayoutSAM, EliGen, and GRIT average only 1.78-3.93 objects per image, limiting supervision for dense multi-object generation.

Key point An automated pipeline keeps only boxes that overlap at least one other box.

The prompt explicitly requires each box to overlap at least one other box and to be at least 5% of image resolution in width and height.

Key point Each object gets a written description over six times longer than in current datasets.

Beyond annotation density, OverLay++ provides rich semantic detail with object captions over six times longer than in current datasets.

Key point The result is roughly 500K images averaging 6.6 objects per image.

OverLay++ contains approximately 500K images with an average of 6.6 objects per image, exceeding existing datasets in annotation density.

Key point Methods trained on the 6.6-object dataset show consistent improvement and faster convergence across benchmarks.

Across multiple benchmarks, state-of-the-art Layout-to-Image methods trained on the dataset show consistent improvement and faster convergence, demonstrating the importance of dense, overlap-aware, and caption-rich supervision for controllable image generation.