UniPercept: Towards Unified Perceptual-Level Image Understanding across Aesthetics, Quality, Structure, and Texture

Multimodal large language models (MLLMs) have achieved remarkable progress in visual understanding tasks such as visual grounding, segmentation, and captioning. However, their ability to perceive perceptual-level image features remains limited. In this work, we present UniPercept-Bench, a unified framework for perceptual-level image understanding across three key domains: Aesthetics, Quality, Structure and Texture. We establish a hierarchical definition system and construct large-scale datasets to evaluate perceptual-level image understanding. Based on this foundation, we develop a strong baseline UniPercept trained via Domain-Adaptive Pre-Training and Task-Aligned RL, enabling robust generalization across both Visual Rating (VR) and Visual Question Answering (VQA) tasks. UniPercept outperforms existing MLLMs on perceptual-level image understanding and can serve as a plug-and-play reward model for text-to-image generation. This work defines perceptual-level image understanding in the era of MLLMs and, through the introduction of a comprehensive benchmark together with a strong baseline, provides a solid foundation for advancing perceptual-level multimodal image understanding.

Today’s multimodal AI systems can often recognize objects, describe scenes, and point to regions in an image, but they still struggle with more subtle visual judgments: whether an image is beautiful, sharp, natural-looking, well-structured, or rich in texture. These qualities matter because they shape how people judge photos, designs, and AI-generated images, yet they are hard to evaluate with existing vision benchmarks. We introduce UniPercept-Bench, a benchmark for measuring this kind of perceptual image understanding across aesthetics, image quality, structure, and texture. To build it, we organize these visual concepts into a clear hierarchy and collect large-scale data for both rating images and answering questions about them. We also train UniPercept, a model designed to perform well across these tasks by learning from perceptual visual data and task-aligned feedback. Experiments show that UniPercept performs better than existing multimodal language models at understanding these perceptual properties. Beyond evaluation, UniPercept can also be plugged into text-to-image systems as a scoring model to help generate images that better match human visual preferences.