Wiki

IQA / IAA Datasets

IQA / IAA Datasets

The fixed coordinate system the whole field reports against. Methods come and go (Image Quality Assessment (IQA), Image Aesthetic Assessment (IAA)); these datasets are what every SRCC/PLCC number (IQA / IAA Evaluation Metrics) is measured on.

The spine

DatasetYearTaskImagesDistortionLabelSubjective study
LIVE IQA (R2)2006IQA779 (+29 ref)synthetic, 5 typesDMOS 0–100~23 subjects, in-lab, single-stimulus w/ hidden ref
CSIQ2010IQA866 (+30 ref)synthetic, 6 typesDMOS 0–1~35 subjects, ~5,000 ratings, in-lab
TID20132013IQA3,000 (25×24×5)synthetic, 24 types × 5MOS 0–9971 observers, ~524k pairwise comparisons
KADID-10k2019IQA10,125 (81×25×5)synthetic, 25 types × 5MOS30 crowd ratings/image
CLIVE (LIVE-in-the-Wild)2016IQA1,162authenticMOS>350k scores, >8,100 MTurk workers
KonIQ-10k2020IQA10,073authenticMOS1.2M ratings, 1,459 crowd workers
SPAQ2020IQA11,125authentic (66 phones)MOS + 5 attributes + EXIFin-lab, controlled
PaQ-2-PiQ / FLIVE2020IQA~39,810 + 120k patchesauthenticMOS (picture + patch)~4M human judgments
AVA2012IAA255,530nonescore distribution (1–10) + 66 tags + 14 stylesDPChallenge, avg ~210 votes/image
AADB2016IAA10,000nonescore + 11 attributes5 MTurk raters/image, identity tracked
PARA2022IAA31,220nonescore + 9 objective + 4 subjective attributes + rater traits438 subjects, ~25 annotations/image
TAD66K2022IAA66,327nonetheme-aware scores, 47 themes>1,200 annotations/image
EVA2020IAA4,070noneMOS + 4 attributes + difficulty + weights≥30 votes/image
FLICKR-AES2017IAA (PIAA)~40,000noneper-rater scores (~200 imgs/worker)210 workers, rater identity tracked

Ref = pristine reference-image count. Study-scale figures are primary-source numbers where stated. Unverified against primary source (protocol confirmed, exact count not stated in abstracts): SPAQ rating count, PARA total annotations, EVA total votes. LIVE’s single- vs double-stimulus wording differs across secondary sources; primary description is single-stimulus with hidden reference. KADID-10k has an unlabelled companion KADIS-700k (700k images) for weak supervision. FLICKR-AES (Ren et al., ICCV 2017) is the first personalized-IAA (Image Aesthetic Assessment (IAA)) benchmark — it tracks rater identity so a model can be evaluated per person; its ~200-images/worker figure is approximate. PARA (2022) is the modern PIAA benchmark.

The three transitions this table encodes

  1. Small synthetic → large synthetic (LIVE/TID → KADID-10k). LIVE’s 779 and TID’s 3,000 images are too small for deep nets; KADID-10k + KADIS-700k feed data-hungry CNNs while keeping the clean synthetic design.
  2. Synthetic → authentic (→ KonIQ, CLIVE, SPAQ, FLIVE). The transition that mattered most: synthetic-trained models did not survive real photos. KonIQ (1.2M ratings) and FLIVE (~4M judgments, patch-level) are the modern NR-IQA proving grounds.
  3. Scalar MOS → score distribution (AVA). AVA’s per-image vote histogram is what enabled NIMA’s EMD distribution prediction; AADB/PARA/EVA add named attributes (explainable aesthetics); TAD66K adds theme awareness.

Why this matters for reading method papers. A 2013 method reporting only LIVE/TID SRCC and a 2023 method reporting KonIQ/FLIVE SRCC are not comparable — they solved different problems. NR-IQA methods after ~2019 benchmark on KonIQ and FLIVE; IAA methods benchmark on AVA.

The MLLM era — instruction & benchmark datasets

A fourth kind of dataset arrived with Multimodal-LLM Visual Scoring (2023–2026): not MOS-regression sets but instruction-tuning corpora (human descriptions of quality/aesthetics, and synthesized question-answer pairs) and capability benchmarks (perception / description / comparison, scored beyond a single correlation). The MOS spine above is still what SRCC/PLCC is measured on; these teach and probe the language side.

DatasetYearForContent
Q-Pathway2023IQA (instruction)58,000 human low-level descriptions on 18,973 images (→ Q-Instruct)
Q-Instruct2023IQA (instruction)200,000 instruction-response pairs synthesized from Q-Pathway
Co-Instruct-562K2024IQA (instruction)multi-image comparative quality instructions
AesMMIT2024IAA (instruction)409K aesthetic instructions over 21,904 images + 88K feedbacks (→ AesExpert)
UNIAA-LLaVA data2024IAA (instruction)aesthetic visual-instruction data unified from existing IAA sets
Q-Eval-100K2025generative100K text-to-image/video instances, 960K human MOS (quality + prompt alignment)
Q-Bench (LLVisionQA / LLDescribe)2024IQA (benchmark)2,990-image perception MCQs; 499-image expert descriptions
MICBench2024IQA (benchmark)first multi-image quality comparison benchmark for LMMs
UNIAA-Bench2024IAA (benchmark)aesthetic perception / description / assessment
Q-Bench-Video2024VQA (benchmark)2,378 QA pairs on video quality understanding

These do not carry a single per-image MOS to regress; they carry descriptions, comparisons, or QA. The convergence-era point (R4): the field is adding datasets that measure *describe / compare / reason about quality, not just predict a number — and generative-image and video quality (Q-Eval-100K, Q-Bench-Video) are folding into the same MLLM-scorer pipeline.*

See also

Trending Tags