Research map / Vision-language and multimodal / Multimodal LLMs

R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment

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

Builds a 3.5K-image computer graphics (CG) quality dataset with six-dimension text descriptions and proposes R4-CGQA, a training-free two-stream (CLIP content + REIQA quality) retrieval-augmented prompting method that improves VLMs on CG quality QA.

Problem

Existing CG quality datasets only provide scalar scores without explanations, and VLMs are inaccurate at fine-grained CG quality judgment and hallucinate in this domain.

Approach

Retrieve example quality descriptions for a query CG image using a fused score of CLIP content similarity and quality-aware (REIQA) embedding similarity, and inject them into the VLM prompt without fine-tuning; the retrieval rule is framed as a MAP estimator with two views.

Key results

• R4-CGQA improves all 11 evaluated VLMs on choice and yes-or-no accuracy; e.g., Bakllava-7B choice accuracy rises from 43.72% to 55.97% (+12.25%).
• Gemma3-4B yes-or-no accuracy improves from 53.55% to 65.22% and Q&A score from 1.05 to 2.32.
• Ablation on LLaVA-1.6-7B: full content+quality retrieval reaches 59.8% choice and 59.9% yes-or-no vs 50.1% and 48.8% base.

Surprising findings

• Feeding multiple similar images directly into Pixtral reduces choice accuracy, so text-description retrieval is preferred over multi-image input.
• Including irrelevant or dissimilar descriptions (T=1.0, no retrieval) sharply degrades yes-or-no performance.

Limitations

• Q&A evaluation relies on GPT-4o-mini as judge rather than human ratings.
• Retrieval depends on a fixed library and similarity threshold/K hyperparameters that must be tuned on validation set.
• Dataset descriptions and validation/test questions are generated by GPT-4o, possibly introducing bias.

Experimental setup

Datasets: CGIQA-6K R4-CGQA base set

Benchmarks: R4-CGQA testing set

Metrics: accuracy (choice) accuracy (yes-or-no) Q&A score (5-point scale)

Baselines: LLaVA-1.6-7B LLaVA-1.6-13B Llama 3.2-Vision-11B Gemma3-4B MiniCPM-V-8B Bakllava-7B Pixtral-12B LlavaNext-8B LlavaNext-32B Qwen2.5-VL-7B Qwen2.5-VL-32B

Builds on: REIQA CLIP FAISS Q-Bench Q-Instruct

Closest papers by meaning

• Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.87)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.86)
• VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations (ICLR 2026 · similarity 0.86)
• Q-Bench-Video: Benchmark the Video Quality Understanding of LMMs (CVPR 2025 · similarity 0.85)
• Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
• Gen3DEval: Using vLLMs for Automatic Evaluation of Generated 3D Objects (CVPR 2025 · similarity 0.85)