Research map / Vision-language and multimodal / Multimodal LLMs

UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 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.

Proposes UARE, a unified Bagel-based vision-language model that performs image quality assessment, restoration, and enhancement in one model via a two-stage (progressive degradation curriculum, then unified IQA-restoration fine-tuning) training framework.

Problem

Image quality assessment and image restoration/enhancement are typically treated in isolation, and no unified model had studied how IQA can guide restoration across multiple degradation types.

Approach

A mixture-of-transformers unified model (built on Bagel) with separate IQA and restoration experts sharing self-attention; restoration is trained with rectified flow after a progressive single-to-mixed-to-high-order degradation curriculum, then jointly fine-tuned with IQA text and interleaved analyze-then-restore data so the generated quality analysis conditions the restoration.

Key results

• On RealSR, UARE reports MUSIQ 69.67 and TOPIQ 0.6796, the best among compared methods, with PSNR 21.38 which is lower than most diffusion baselines.
• On the KADID and CSIQ IQA benchmarks, UARE reaches PLCC 0.878 and 0.930, surpassing Q-Insight (0.757 and 0.768) on those sets.
• In a user study with 29 experts, UARE received 86.21% of votes against all-in-one restoration baselines and 64.14% against SR baselines.

Surprising findings

• The PSNR of UARE on RealSR (21.38) and DRealSR (21.31) is much lower than most diffusion SR baselines, trading fidelity for perceptual metrics.
• In the IQA benchmark, UARE underperforms Q-Insight on SPAQ/LIVEW/AGIQA in several metrics, and its AGIQA SRCC (0.667) is far below baselines.

Limitations

• Large model size and slow inference limit deployment on resource-constrained devices.
• Restoration and enhancement improving IQA is not explored.
• Lower PSNR than several SR baselines on RealSR and DRealSR.

Experimental setup

Datasets: KONIQ KADID PIPAL SPAQ DQ-495K DiffIQA FoundIR LSDIR RealESRGAN degradation pipeline APISR degradation pipeline

Benchmarks: RealSR DRealSR DIV2K RealSet80 FoundIR SPAQ KADID LIVEW AGIQA CSIQ

Metrics: PSNR SSIM LPIPS DISTS NIQE LIQE MUSIQ MANIQA TOPIQ PLCC

Baselines: StableSR DiffBIR SeeSR PASD OSEDiff S3Diff PURE Q-Insight Q-Align DeQA-Score FoundIR SUPIR

Builds on: Bagel Q-Insight PURE rectified flow FoundIR

Closest papers by meaning

• LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models (NeurIPS 2026 · similarity 0.89)
• IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.89)
• Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.88)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.88)
• Evaluation of Visual Processing Should Be Human-Centered, Not Metric-Centered (NeurIPS 2026 · similarity 0.88)
• Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.88)