ABACUS model masters image counting and generation with 3B parameters

ABACUS, developed by researchers Anindya Mondal, Sauradip Nag, and Anjan Dutta, is a unified vision-language model that tackles multiple counting tasks—object counting, crowd counting, and referencing-expression counting—alongside count-faithful image generation, all without requiring benchmark-specific fine-tuning. Built on an existing 3B-parameter foundation model, it introduces three key innovations: density-aware adaptive zooming with objectness maps for spatial grounding, a boundary-aware count policy via GRPO (Group Relative Policy Optimization) to eliminate crop-boundary errors, and a cycle-consistent GRPO strategy where the understanding branch critiques generated outputs, closing the understanding-generation gap without external annotations.

This approach achieves state-of-the-art results across seven benchmarks, outperforming both task-specific specialists and larger generalist models. The cycle-consistent GRPO is particularly notable as it enables the model to self-correct its generated images based on counting accuracy, a capability that previously required separate pipelines or annotation-heavy training. The paper is currently under review and available on arXiv. For professionals in computer vision, ABACUS represents a significant step toward truly unified models that can both understand and generate visual content with precise numerical fidelity.