ABACUS: Adapting Unified Foundation Model for Bridging Image Count Understanding and Generation

2026-06-22

AI Summary

This work addresses the longstanding disconnect between existing counting tasks—such as object counting, crowd counting, and referential expression counting—and count-preserving image generation, which have traditionally required separate modeling approaches. The authors propose a unified framework based on a 3B-parameter vision-language foundation model that jointly handles diverse counting and generation tasks without task-specific training. Key innovations include density-aware adaptive scaling and objectness maps for precise spatial localization, a boundary-aware GRPO strategy to mitigate cropping artifacts, and a generative reinforcement learning mechanism enforcing cycle consistency between understanding and generation. Evaluated across seven benchmarks, the method outperforms both specialized models and larger general-purpose counterparts, achieving significant gains in accuracy and generalization.

Abstract

ABACUS is a unified vision-language model that handles object counting, crowd counting, referring-expression counting, and count-faithful image generation without any benchmark-specific training required. Our model is built on existing 3B-parameter unified foundation model and is adapted for object localization tasks using three key innovations: density-aware adaptive zooming with objectness maps for spatial grounding; a boundary-aware count policy via GRPO to eliminate crop-boundary errors; and a cycle-consistent GRPO strategy where the understanding branch self-critiques generated outputs, closing the understanding-generation gap without any external annotations. ABACUS achieves state-of-the-art results across seven benchmarks, outperforming both task-specific specialists and larger generalist models.

Problem

• object counting
• crowd counting
• referring-expression counting
• count-faithful image generation
• unified vision-language model

Innovation

• density-aware adaptive zooming
• boundary-aware count policy
• cycle-consistent GRPO
• count-faithful generation
• unified vision-language model