Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces UniPercept-Bench, a hierarchical benchmark for perceptual-level image understanding (aesthetics, quality, structure/texture) with VR and VQA tasks, and UniPercept, an InternVL3-8B-based model trained via domain-adaptive pretraining and GRPO with soft rewards, usable as a T2I reward model.

Problem

Multimodal LLMs struggle to perceive and evaluate perceptual-level image attributes (aesthetics, quality, structure and texture), which lack standardized definitions, benchmarks, and strong models.

Approach

A three-level Domain-Category-Criterion taxonomy for perceptual attributes including a new ISTA (structure and texture) domain, paired with a two-stage training recipe (domain-adaptive pretraining then task-aligned GRPO with an adaptive Gaussian soft reward for ratings and binary reward for VQA).

Key results

• UniPercept achieves average SRCC/PLCC of 0.590/0.586 on UniPercept-Bench VR-IAA, outperforming the best listed baseline (0.431/0.410 GPT-4o among general models, 0.398/0.395 ArtiMuse).
• UniPercept reaches 76.55% overall accuracy on UniPercept-Bench VQA-IAA and 84.23% on VQA-ISTA, above the best general model (GLM/InternVL3-78B-level, about 68%) on IAA.
• As a reward for FLUX.1-dev via Flow-GRPO, the UniPercept IAA reward raises ArtiMuse score from 59.02 to 67.02 and UniPercept IAA score from 65.18 to 76.20.

Surprising findings

• Many general-purpose MLLMs, including GPT-4o and Gemini-2.5-pro, score near zero or negative correlation on ISTA rating, indicating poor direct numerical perception.
• Scores from the authors' own ISTA-10K rating are rule-based counts over generated structured annotations rather than human judgments.

Limitations

• UniPercept-Bench is smaller than typical semantic-level benchmarks (about 6K items).
• ISTA rating relies on a deterministic lexicon-based scoring rule and GPT-4o-generated annotations, which may not reflect human judgments.
• Dataset-level and T2I model distribution analyses are preliminary and left for future work.

Experimental setup

Datasets: ArtiMuse-10K KonIQ-10K ISTA-10K APDDv2 Impressions Q-Ground-100K DataDepictQA DTD FMD Flickr2K LSDIR AVA

Benchmarks: UniPercept-Bench-VR UniPercept-Bench-VQA AVA TAD66K FLICKR-AES KonIQ-10K SPAQ KADID PIPAL DPG GenEval

Metrics: SRCC PLCC accuracy (SRCC+PLCC)/2 DPG overall GenEval overall

Baselines: GPT-4o Gemini-2.5-pro Claude-Sonnet-4.5 Llama-4-Scout InternVL3-78B QwenVL-2.5-Instruct-72B Q-Align ArtiMuse DeQA Q-Insight

Builds on: GRPO Q-Align Flow-GRPO Token As Score

Closest papers by meaning

• UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning (NeurIPS 2025 · similarity 0.89)
• UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.89)
• JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.89)
• UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation (NeurIPS 2025 · similarity 0.89)
• Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.88)
• Hugging Visual Prompt and Segmentation Tokens: Consistency Learning for Fine-Grained Visual Understanding in MLLMs (CVPR 2026 · similarity 0.88)