Learning to Generate Multiple Objects from Dense and Occluded Layouts

Bach-Hoang Ngo1,2,†   Si-Tri Ngo1,2,†   Hieu Le3,‡   Trung-Nghia Le1,2,‡,*

† Equal contribution ‡ Equal advising * Corresponding author

1University of Science, Ho Chi Minh City   2Vietnam National University, Ho Chi Minh City   3UNC Charlotte

Project Page
Count-preserving multi-category compositional generation results
Count-preserving multi-category compositional generation. Standard diffusion models fail as scene complexity grows, while our approach maintains more accurate counts under dense layouts and occlusion.

TL;DR: AIBL improves multi-object generation by combining ownership-aware layout attention with an amodal-aware instance-balanced loss for occluded objects.

Abstract

Text-to-image diffusion models often fail to produce the requested number of objects in dense scenes. Overlapping instances can collapse into merged structures, and heavily occluded objects receive weak training signal because only small visible regions contribute to the loss.

We introduce AIBL, a layout-aware training framework that encourages instance ownership throughout generation. Ownership-aware attention biases reduce cross-instance feature leakage, while an amodal-aware instance-balanced loss increases supervision for occluded objects according to their amodal-to-visible area ratio.

Method Overview

AIBL architecture overview
Architecture overview. Layout Attention steers generation with region identity maps, while the AIBL training objective prioritizes heavily occluded instances using amodal and visible masks.

Qualitative Results

Qualitative comparison on counting prompts
Qualitative comparison on counting prompts. AIBL promotes clearer instance separation compared with layout-only generation and contemporary text-to-image baselines.

Analysis

Counting score stratified by target object density
Counting score stratified by target object density.
Gamma ablation for AIBL
Sensitivity of AIBL to the dampening exponent.

Failure Cases

AIBL failure cases
Failure cases. The method can still struggle with extreme density or thin objects.

Resources

This is a temporary public project page. Source code and dataset links are coming soon.

Source Code

Training, inference, and evaluation scripts will be linked here when ready.

Dataset

OverlapDepth access and documentation will be linked here when released.

Project URL

bachngoh.github.io/AIBL

BibTeX

@misc{ngo2026aibl,
  title  = {Learning to Generate Multiple Objects from Dense and Occluded Layouts},
  author = {Ngo, Bach-Hoang and Ngo, Si-Tri and Le, Hieu and Le, Trung-Nghia},
  year   = {2026},
  note   = {Project page. Paper, code, and dataset coming soon}
}