Research map / Vision-language and multimodal / Multimodal LLMs

Probabilistic Prompt Adaptation for Unified Image Aesthetics and Quality Assessment

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

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 Probabilistic Prompt Adaptation (PPA), which marginalizes over LLM-pre-sampled text prompts with a learned image-task prompt-selection model on top of CLIP-IQA, enabling annotation-free, prompt-controllable IAA/IQA scoring.

Problem

Existing VLM-based IAA/IQA models either support arbitrary text prompts with limited accuracy (e.g., CLIP-IQA) or achieve high accuracy only for fixed criteria/datasets, lacking flexible prompt-controllable scoring.

Approach

Treat text prompts as latent variables: a prompt-selection model p(t|x,c) weights LLM-sampled antonymous prompts per image and task, and the score distribution is a mixture over prompts trained by marginal likelihood using only (task, image, score) triplets.

Key results

• PPA achieves SOTA on PARA (SRCC 0.497), BAID-free comparisons aside, and SPAQ (SRCC 0.934) and reaches best SRCC/PLCC on PARA, BAID and SPAQ among compared methods; PPA-T gets 0.941 SRCC on SPAQ.
• In a 757-participant crowdsourced study, PPA was preferred over CLIP-IQA for most prompts (e.g., 200/100 for 'focused high-quality photo') and over UniQA for most prompts, with several losses.
• Ablation: task- and image-conditioned prompt weighting raises average SRCC from 0.759 (uniform) to 0.808 across 12 datasets.

Surprising findings

• Fixed-prompt fine-tuning and PPA both improve feature separability, but the paper reports that PPA's prompt-level consistency drops for abstract or emotional aesthetic prompts.

Limitations

• Consistency decreases for abstract or high-level aesthetic prompts, attributed to limited prompt diversity.
• Prompt pool is fixed to LLM-generated antonymous pairs of 260 prompts.
• PPA trails best methods on several IAA datasets (e.g., AVA, AADB) by a small margin.

Experimental setup

Datasets: AVA AADB TAD66k PARA BAID KonIQ-10k SPAQ TID2013 KADID-10k CSIQ LIVE LIVEC

Benchmarks: AVA AADB TAD66k PARA BAID KonIQ-10k SPAQ TID2013 KADID-10k CSIQ LIVE LIVEC

Metrics: SRCC PLCC silhouette score between/within-class variance ratio human win ratio

Baselines: CLIP-IQA UniQA NIMA MUSIQ Charm IAACLIP Gamma Gamma-T Gamma+ LIQE GRMP-IQA GIAA

Builds on: CLIP-IQA CoOp UniQA

Closest papers by meaning

• FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges (NeurIPS 2025 · similarity 0.86)
• DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models (NeurIPS 2026 · similarity 0.86)
• IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.85)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.85)
• Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.85)
• Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.85)