Скачать или смотреть UniPercept: Perceptual Image Benchmark

In this AI Research Roundup episode, Alex discusses the paper: 'UniPercept: Towards Unified Perceptual-Level Image Understanding across Aesthetics, Quality, Structure, and Texture' UniPercept tackles a key weakness in multimodal large language models: perceptual-level understanding of how images look and feel, beyond semantic captioning and grounding. The paper unifies three domains—Image Aesthetics Assessment, Image Quality Assessment, and Image Structure and Texture Assessment—under a hierarchical Domain–Category–Criterion taxonomy. It introduces UniPercept-Bench, a multi-task benchmark with two evaluation formats: Visual Rating for continuous scoring and Visual Question Answering for multiple-choice and explanation consistency. The dataset is built with a rigorous pipeline combining generator MLLM annotation, automated reject sampling by a judge MLLM, and final human refinement, with extra structure and deterministic scoring for the structure-and-texture domain. The authors also present UniPercept, a baseline MLLM trained from an InternVL3-8B backbone to better align models with human perceptual judgments. Paper URL: https://arxiv.org/abs/2512.21675 #AI #MachineLearning #DeepLearning #MultimodalLLM #ImageAesthetics #ImageQualityAssessment #ComputerVision #Benchmarks

Resources:
GitHub: https://github.com/thunderbolt215/Uni...
Hugging Face model: https://huggingface.co/Thunderbolt215...