Wiki

Multimodal-LLM Visual Scoring

Multimodal-LLM Visual Scoring

The 2023–2026 convergence era of IQA and IAA: the field stopped training a bespoke regressor per task and started teaching a pretrained multimodal large language model (MLLM / LMM) to rate. The result is one model that scores technical quality, aesthetics, and video quality at or above task-specific SOTA — R1’s “same MOS-prediction machine, opposite questions” thesis made literal. This page holds the material that spans both tracks; the IQA- and IAA-specific rungs below it live in Image Quality Assessment (IQA) and Image Aesthetic Assessment (IAA). Full detail is survey report R4.

The pivot

A pretrained MLLM already carries a rich, general prior over what images are and what people say about them (from web-scale image-text). So the task shrinks from “train a quality model” to “elicit and align the opinion the model already has.” This is the terminus of R2’s CLIP on-ramp (CLIP-IQA, LIQE) and R3’s VILA on-ramp — and it generalizes both.

It dissolves both frontier problems with one instrument — MLLM world-knowledge plus language grounding:

  • IQA’s cross-dataset generalization gap (Image Quality Assessment (IQA), R2): the external prior is not perturbed by any single dataset’s camera/rater population, so it transfers. Q-Align holds KonIQ→CLIVE 0.860 where HyperIQA fell to 0.785.
  • IAA’s underdetermined scalar / subjectivity (Image Aesthetic Assessment (IAA), R3): the MLLM’s native output is language, so it can hold attributes, taste, and rater disagreement a scalar erased — then collapse them to a number if wanted.

The qualitative leap is explainability: a score and a natural-language critique from one forward pass — auditable quality, defensible aesthetics.

The levels-as-tokens trick (Q-Align)

The key mechanism, and why it works. Instead of regressing a number:

  1. Reframe the label as five discrete words. Human studies (IQA / IAA Evaluation Metrics ITU protocols) asked raters for a category, not a real number. Q-Align teaches the LMM exactly excellent, good, fair, poor, bad → ${5,4,3,2,1}$. The training target is the word, so the loss lives in the model’s native next-token prediction — no regression head.
  2. Convert the level distribution to a continuous score at inference. Softmax over just the five level-token logits, then a probability-weighted average (Q-Align Eq. 4):

    \[S = \sum_{i=1}^{5} p_{\ell_i}\cdot i,\qquad p_{\ell_i} = \frac{e^{\,x_{\ell_i}}}{\sum_{j=1}^{5} e^{\,x_{\ell_j}}}\]

Why discrete words beat regressing a numeral directly: an LLM is a token predictor, not a function approximator — ordinal adjectives are in-distribution, a continuous numeral is not (Q-Align Table 1: 96–100 % of LMMs spontaneously answer a quality question with a word, not a number). It also matches how the labels were made (MOS = average of categories), and it generalizes far better — the discrete-level syllabus beats direct score regression by +52.8 % SRCC on a SPAQ→KADID cross-dataset transfer (Q-Align Table 11). This is the mechanism that dents R2’s generalization frontier.

The Q-family spine (Q-Future · NTU / SJTU, Wu et al.)

Lead author Haoning Wu; senior authors Weisi Lin, Guangtao Zhai. The arc runs benchmark → describe → rate.

  1. Q-Bench (ICLR 2024 Spotlight, arXiv:2309.14181) — the diagnosis. Three axes: perception (LLVisionQA, 2,990 images), description (LLDescribe, 499 images), assessment. Finding: general MLLMs (GPT-4V, Gemini) have “preliminary low-level visual skills” — they perceive/describe quality above chance — but are “unstable and imprecise” and cannot produce precise quantitative scores. That gap is the era’s motivation.
  2. Q-Instruct (CVPR 2024, arXiv:2311.06783) — teach the model to describe quality. Q-Pathway = 58,000 human low-level descriptions on 18,973 images; Q-Instruct = 200,000 instruction-response pairs synthesized from it. Lifts perception/description, but not yet calibrated scoring.
  3. Q-Align / OneAlign (ICML 2024, arXiv:2312.17090) — the convergence result, built on mPLUG-Owl2. The levels-as-tokens trick above. OneAlign: one model jointly trained on IQA + IAA + video VQA that beats task-specific SOTA on all three, incl. cross-dataset.
  4. Co-Instruct (ECCV 2024 Oral, arXiv:2402.16641) — open-ended multi-image quality comparison (Co-Instruct-562K data, MICBench benchmark). R1’s pairwise-comparison protocol reborn as an LMM capability.

DepictQA (You, Xue, Dong et al., CUHK / XPixel — a different group; ECCV 2024, arXiv:2312.08962) — describe-and-compare quality in language as a hierarchical descriptive judgment; follow-up DepictQA-Wild. The score is a lossy projection of a richer linguistic verdict.

The aesthetic arm — same recipe

Aesthetics lands on the identical instruction-tuning recipe as quality (R1’s “same machine” holding on the aesthetic side):

  • Q-Align’s aesthetic arm — the same discrete-level method on AVA gives 0.822 / 0.817 SRCC/PLCC (top of R3’s AVA ladder) with no aesthetic-specific architecture.
  • UNIAA (Kuaishou/Kling + PKU, arXiv:2404.09619) — a unified aesthetic baseline (UNIAA-LLaVA) + benchmark (UNIAA-Bench, three levels echoing Q-Bench: perception/description/assessment).
  • AesExpert (Huang, Li, Lin, Shi et al., Xidiannot the TANet/TAD66K group; ACM MM 2024, arXiv:2404.09624) — aesthetics instruction tuning at scale: AesMMIT (409K instructions, 21,904 images, 88K feedbacks) → the AesExpert model. Q-Instruct’s move pointed at composition, colour harmony, mood.

Recipe, both tracks: benchmark native ability (Q-Bench / UNIAA-Bench) → instruction-tune to describe (Q-Instruct / AesExpert) → align to rate (Q-Align, both arms).

Benchmark: one model at the top of both columns

In-domain SRCC / PLCC (IQA / IAA Evaluation Metrics). The bottom block is the convergence: one family tops the IQA column R2 built and the AVA column R3 built.

MethodTrackKonIQAVA
LIQE (CLIP, R2)IQA0.919 / 0.9120.776
VILA-R (VLP, R3)IAA0.774 / 0.774
Q-Align (MLLM)both0.940 / 0.9410.822 / 0.817
OneAlign (unified)both + video0.941 / 0.9500.823 / 0.819
DeQA-Score (2025)IQA0.941 / 0.953

OneAlign also scores video (LSVQ 0.886 / 0.886). Cross-dataset (train KonIQ): Q-Align → CLIVE 0.860, SPAQ 0.887 vs HyperIQA→CLIVE 0.785, CLIP-IQA+→CLIVE 0.805 — the smallest generalization collapse in the series. Not attributed to Q-Align: FLIVE (not in its tables); MANIQA (not a Q-Align row — its named IQA baseline is CLIP-IQA+).

Open problems (the honest close)

  • Calibration / reproducibility. A softmax over five tokens is prompt-, temperature-, and checkpoint-sensitive. DeQA-Score (CVPR 2025, arXiv:2501.11561) models the whole score distribution (Thurstone fidelity loss) to address it.
  • Hallucination in the language explanation — a fluent, confident, wrong reason invites misplaced trust.
  • Benchmark saturation + contamination. KonIQ ≈ 0.94 / AVA ≈ 0.82 are near the label noise ceiling; web-scale pretraining risks train/test leakage, so a high SRCC may measure memorization.
  • Compute. OneAlign is billions of parameters vs MANIQA’s ≈ 20 MB / NIQE’s millisecond CPU cost — a 1000× gap for ≈ 0.02–0.05 SRCC.
  • Is SRCC-on-one-dataset the right target anymore? If the model can describe, compare, and reason (Q-Bench, Co-Instruct, DepictQA), one correlation coefficient throws away most of what it does — the frontier may be faithful description/comparison, not a number.

Third axis folding in. R1 kept generative-image quality (“how real?”) and video quality (VQA) separate; both are being absorbed: OneAlign scores video; Q-Bench-Video (CVPR 2025, arXiv:2409.20063) benchmarks LMM video understanding; Q-Eval-100K / Q-Eval-Score (CVPR 2025 Oral, arXiv:2503.02357) scores generated content quality + prompt alignment. R1’s three axes are converging on one linguistic scorer.

Q-Insight (arXiv:2503.22679, 2025) carries the “Q-“ name but is a different group (Jian Zhang et al., not Q-Future) — an RL/reasoning-based quality MLLM. Do not attribute it to Haoning Wu’s cluster.

Practitioner decision guide

  • ≥ 0.9 SRCC + interpretability, compute available: Q-Align / OneAlign (or DeQA-Score for calibration).
  • Tiny / fast / on-device: MANIQA or HyperIQA (≈ 0.90, few MB); NIQE for zero-training millisecond CPU scoring (accept ≈ 0.66 authentic).
  • Perceptual training loss (SR / diffusion / codecs): LPIPS or DISTS (R2’s FR branch) — MLLM scorers are not differentiable losses.
  • Compare images / need a reason, not a number: Co-Instruct or DepictQA.

See also

Trending Tags