Image Quality and Aesthetic Assessment Converge: Multimodal LLMs as Unified Visual Scorers

The fourth and final report of a four-report survey series building a domain mental model of Image Quality Assessment (IQA) and Image Aesthetic Assessment (IAA). R1 mapped the problem space; R2 walked the IQA method lineage to its CLIP on-ramp; R3 walked the IAA lineage to its VILA on-ramp. This report is the convergence finale: the 2023–2026 era in which the field stopped training a bespoke regressor per task and started teaching a pretrained multimodal large language model (MLLM / LMM) to rate — and IQA, IAA, and even video quality collapsed into one model. It mirrors R2/R3’s house style — a transition-labelled lineage diagram, a consolidated benchmark table, and “what broke → what fixed it” prose — and it closes the loop back to R1’s thesis that IQA and IAA are the same MOS-prediction machine pointed at opposite questions. R4’s claim is that the machine literally became one model. It assumes the whole series’ vocabulary (FR/NR, KonIQ/CLIVE/AVA, SRCC/PLCC, the five-era arc) and does not re-explain it.

Short answer

The convergence is real and it is one model. Q-Align’s OneAlign — a single multimodal LLM — sits at or above the top of both leaderboards at once: ≈ 0.94 SRCC on KonIQ-10k (technical quality, beating the best CLIP-era IQA regressor) and ≈ 0.82 SRCC on AVA (aesthetics, beating VILA), while the same weights also score video quality (LSVQ ≈ 0.89). R1’s dotted line — “same machine, opposite questions” — stopped being an analogy and became a checkpoint file.

The pivot that made it happen, stated as R2/R3’s failure/repair chain:

1. The reframe. Instead of fitting a numeric regressor per dataset, you take a pretrained MLLM (Q-Align uses mPLUG-Owl2) and teach it to rate the way humans were taught to rate: with the five discrete words bad, poor, fair, good, excellent. At inference you take the softmax over just those five level-token logits and read a continuous score as their ${1,2,3,4,5}$-weighted average. Why discrete words, not a number: an LLM is a next-token predictor, and 96–100 % of LMMs spontaneously answer “how is the quality?” with an adjective, not a numeral (Q-Align, Table 1); ordinal words are in-distribution for it, a continuous numeral is not — the discrete-level syllabus beats direct score regression by +52.8 % on a cross-dataset transfer in Q-Align’s own ablation. This is the key mechanism; §The levels-as-tokens trick gives it room.

2. Why it dissolves both frontier problems at once. R2’s open problem was cross-dataset generalization (in-domain ≈ 0.92 collapsing to ≈ 0.75); R3’s was subjectivity and the underdetermined scalar (beauty is a distribution over disagreeing people that a number cannot hold). A pretrained MLLM attacks both with the same asset — world knowledge + language grounding. World knowledge is what lets Q-Align cross KonIQ→CLIVE at 0.860 where HyperIQA fell to 0.785: it did not overfit one dataset’s camera population, it knows what a good photo is. And because it reasons in language about why, it can express the taste, uncertainty, and attributes a scalar erased — which is IAA’s whole problem.

3. The qualitative leap is explainability. The output is no longer a bare number but a score plus a natural-language critique (“slightly underexposed, soft focus on the subject, pleasing warm palette”). Q-Bench showed general MLLMs (GPT-4V) already perceive low-level attributes but score imprecisely; Q-Instruct taught them to describe quality; Q-Align taught them to rate it — and Co-Instruct / DepictQA push on to compare and depict quality in open language. Aesthetics followed the identical recipe (UNIAA, AesExpert). The number and the reason now come from one forward pass.

The honest caveat, carried to the close: these gains cost a multi-billion- parameter model against MANIQA’s ~20 MB, MLLM scores raise calibration and reproducibility questions, the language explanations can hallucinate, and web-scale pretraining makes benchmark contamination a live worry — so whether “SRCC on KonIQ/AVA” is even the right target anymore is itself now in play (§Frontier). The practitioner’s decision guide is in §Series synthesis.

The convergence in one picture

The two lineages R2 and R3 tracked separately run into the same node. Read it as R2’s IQA relay (top) and R3’s IAA relay (bottom) climbing in parallel, each arriving at a vision-language on-ramp (LIQE / VILA), then merging into the MLLM-scoring era where one model answers both — and video too.

graph LR
  subgraph IQAtrack["IQA track (R2)"]
    Q1["NSS<br/>BRISQUE·NIQE"] -->|learn features| Q2["CNN<br/>DBCNN·HyperIQA"]
    Q2 -->|native-res attn| Q3["ViT<br/>MUSIQ·MANIQA"]
    Q3 -->|prompt a VLM| Q4["CLIP<br/><b>CLIP-IQA·LIQE</b>"]
  end
  subgraph IAAtrack["IAA track (R3)"]
    A1["Rules<br/>Datta·Ke"] -->|learn + distribution| A2["CNN<br/>NIMA·MLSP"]
    A2 -->|attributes·theme| A3["Attr/Theme<br/>AADB·TANet"]
    A3 -->|read comments| A4["VLP<br/><b>VILA</b>"]
  end
  Q4 -->|"open problem:<br/>cross-dataset generalization"| M
  A4 -->|"open problem:<br/>subjectivity / underdetermined scalar"| M
  M["<b>MLLM scoring · 2023–2026</b><br/>Q-Bench → Q-Instruct → <b>Q-Align / OneAlign</b><br/>Co-Instruct · DepictQA · UNIAA · AesExpert<br/><i>teach a pretrained LMM discrete rating levels →<br/>softmax over level-tokens = one continuous score</i>"]
  M -->|"one model, both tracks + video"| OUT["OneAlign<br/>KonIQ ≈0.94 · AVA ≈0.82 · LSVQ ≈0.89"]

The structural claim of the diagram: the two on-ramps are not two branches of one MLLM family — they are the same family entered from two sides. Q-Align is the merge node, and its unified model OneAlign is the single artifact both R2 and R3 were climbing toward without knowing it.

graph LR
  subgraph Spine["The Q-family spine (Q-Future / NTU · SJTU, Wu et al.)"]
    B["<b>Q-Bench</b> '24 ICLR<br/>benchmark: MLLMs<br/><i>perceive</i> low-level,<br/><i>score</i> imprecisely"]
    -->|"fix: teach them to<br/>describe quality"| I["<b>Q-Instruct</b> '24 CVPR<br/>Q-Pathway (58k feedbacks)<br/>→ 200k instruction pairs"]
    -->|"fix: teach discrete<br/>rating levels → score"| AL["<b>Q-Align / OneAlign</b> '24 ICML<br/>IQA + IAA + VQA in one model,<br/>beats task-specific SOTA"]
    -->|"describe & compare<br/>in open language"| CO["<b>Co-Instruct</b> (multi-image)<br/><b>DepictQA</b> (depict + compare)"]
  end

The pivot thesis: teach a pretrained MLLM to rate

Every era in R2 and R3 trained a dedicated model for the quality function — BRISQUE’s SVR, HyperIQA’s hyper-network, MUSIQ’s transformer, even LIQE’s CLIP head. The reference to “what a good image looks like” had to be installed into that model, either by hand (NSS priors) or by ImageNet transfer plus MOS fine-tuning. The convergence era’s single idea is to stop installing it and start borrowing it: a multimodal LLM pretrained on web-scale image-text already carries a rich, general prior over what images are and what people say about them, so the task shrinks from “train a quality model” to “elicit and align the quality opinion the model already has.”

This is the natural terminus of R2’s CLIP on-ramp and R3’s VILA on-ramp, and it generalizes both. CLIP-IQA proved a two-word prompt (“Good photo” vs “Bad photo”) extracts a usable quality prior with zero labels; VILA proved aesthetic knowledge is carried in the comments people wrote. The MLLM move asks the obvious next question: why stop at two prompts or a frozen dual encoder — why not let a full language model look at the image, describe it, reason about it, and rate it?

Why this attacks R2’s and R3’s opposite frontier problems with one instrument:

• Against IQA’s cross-dataset gap (R2). A regressor trained on KonIQ partly fits KonIQ’s camera population and rater pool, so it transfers poorly. An MLLM brings an enormous external prior that no single IQA dataset can perturb — it already knows what noise, blur, and good exposure are across the entire visual web — so it depends far less on the training set’s idiosyncrasies. The cross-dataset numbers in §Consolidated benchmark bear this out directly.
• Against IAA’s underdetermined scalar (R3). Beauty is linguistic all along (R3’s close): the ground truth was never really a number, it was ~210 opinions and the comments beneath each photo. An MLLM’s native output is language, so it can hold the attributes, the taste, and the disagreement a scalar threw away — and then, if you still want a number, collapse its rating-level distribution into one. The scalar becomes a view of a linguistic judgment rather than the judgment itself.

The qualitative leap over everything in R2/R3 is explainability: a score and a reason, from one model. That is not a cosmetic bonus — it is what makes the score auditable (you can see why it said 3.2) and what makes the aesthetic score defensible (the “why” is exactly the content R3’s attribute schemas were groping toward, now open-ended).

The Q-family: the spine of the convergence

The through-line of this era is one research cluster — the Q-Future group (Nanyang Technological University and Shanghai Jiao Tong University), lead author Haoning Wu, senior authors Weisi Lin and Guangtao Zhai. Their three-paper arc is the convergence, and it runs benchmark → describe → rate.

Q-Bench — the diagnosis (ICLR 2024, Spotlight)

Q-Bench asked a prior question: before building an MLLM scorer, what can general-purpose MLLMs already do on low-level vision? It benchmarks GPT-4V, Gemini, Qwen-VL-Plus and ~16 open models on three axes:

1. Perception — the LLVisionQA set (2,990 images, quality-attribute multiple-choice questions: is it blurry, noisy, well-lit?).
2. Description — the LLDescribe set (expert low-level descriptions on 499 images), scoring how well an MLLM narrates quality.
3. Assessment — a softmax-over-tokens strategy to extract a predictable quality score from the MLLM on standard IQA datasets.

The load-bearing finding, and the one that set up the whole era: general MLLMs have “preliminary low-level visual skills” — they perceive and describe quality attributes well above chance — but those skills are “unstable and relatively imprecise”, and the models cannot produce precise quantitative scores. In R1’s terms: MLLMs came pre-loaded with the low-level perception IQA needs, but not the calibrated scoring. That precise gap is what the next two papers close.

Q-Instruct — teach the model to describe quality (CVPR 2024)

Q-Instruct is the instruction-tuning stage. Its data pipeline is two datasets:

• Q-Pathway — 58,000 human low-level feedbacks on 18,973 images: raw subjective descriptions (“the image is slightly dark with visible noise in the shadows, and the subject is out of focus”), not scores. This is the low-level-vision analog of VILA’s comments (R3), collected on purpose.
• Q-Instruct — 200,000 instruction-response pairs synthesized from Q-Pathway (GPT-assisted), turning descriptions into the question-answer format MLLM fine-tuning consumes.

Fine-tuning open MLLMs on it lifts their low-level perception and description consistently. Q-Instruct teaches the vocabulary and attention of quality — the model learns to talk about clarity, exposure, noise, and composition on demand. What it does not yet do is emit a calibrated, correlation-grade number. That is Q-Align.

Q-Align / OneAlign — the convergence result (ICML 2024)

Q-Align is the paper this whole series was pointing at. It is built on mPLUG-Owl2, and it produces OneAlign: a single model trained jointly on IQA + IAA + video VQA that beats task-specific SOTA on all three, including cross-dataset. The mechanism is worth its own section.

The levels-as-tokens trick — and why regressing a number fails

The core trick is to not regress a number at all. Instead:

1. Reframe the label as five discrete words. Human subjective studies (R1’s ITU protocols) never asked raters for a real number — they asked for one of five categories: excellent, good, fair, poor, bad. Q-Align teaches the LMM exactly those words, mapping them to ${5,4,3,2,1}$ (excellent = 5 … bad = 1). During training the target is the word, so the loss lives entirely in the model’s native next-token prediction — no bolted-on regression head.

2. Convert the level-token distribution to a continuous score at inference. Take the model’s logits, restrict the softmax to just the five level tokens, and read the score as the probability-weighted average of their values (the paper’s Eq. 4):

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

   where $x_{\ell_i}$ is the logit of level word $\ell_i$ and its value is $i \in {1,\dots,5}$. A confident “excellent” gives ≈ 5; a genuine tie between “good” and “fair” gives ≈ 3.5 — the soft mass over ordinal words recovers a continuous, well-behaved score, and it inherits the ordinal structure for free (adjacent words are adjacent numbers).

Why the discrete-word detour beats regressing a numeral directly — three reasons, and the third is the killer:

• An LLM is a token predictor, not a function approximator. Continuous numerals (“3.7”) are tokenized awkwardly and lie off the manifold of natural answers; ordinal adjectives are squarely in distribution. Q-Align (Table 1) measures this: 96–100 % of LMMs spontaneously answer a quality question with a qualitative word, not a number — so the model wants to say “good”, and the method leans into that instead of fighting it.
• It matches how the labels were made. Raters emitted categories; MOS is the average of categories. Predicting the category distribution and averaging it mirrors the data-generating process, so the target is aligned with the human protocol rather than an artifact of it.
• It generalizes far better. The paper’s ablation (Table 11) shows the discrete-level syllabus beats a direct score-regression variant by +52.8 % SRCC on a SPAQ→KADID cross-dataset transfer — the discrete words carry meaning that transports across datasets where a fitted numeric scale does not. This is precisely the mechanism that dents R2’s generalization frontier.

OneAlign — one model, three tasks

Trained with this one syllabus on a mixture of KonIQ (IQA), AVA (IAA), and LSVQ (video VQA), OneAlign is a single set of weights that tops all three. The headline in-domain numbers (verified against the paper’s tables):

• IQA: KonIQ 0.941 / 0.950 SRCC/PLCC, SPAQ 0.932 / 0.935, KADID 0.941 / 0.942.
• IAA: AVA 0.823 / 0.819.
• Video: LSVQ_test 0.886 / 0.886, KoNViD-1k (cross) 0.876 / 0.888.

The task-specific Q-Align variants match or exceed the best of R2 and R3 head to head: on IQA it beats CLIP-IQA+ (KonIQ 0.895 → 0.940; cross KonIQ→CLIVE 0.805 → 0.860), and on aesthetics it beats VILA (AVA 0.774 → 0.822). One model, both of the series’ leaderboards, at the top.

Co-Instruct and DepictQA — the comparative / descriptive frontier

Q-Align emits a score with an implicit critique. The next question is whether an MLLM can do the genuinely linguistic quality tasks a regressor never could:

• Co-Instruct (Q-Future, ECCV 2024 Oral) — open-ended, multi-image quality comparison. It is trained on Co-Instruct-562K and evaluated on MICBench (the first multi-image comparison benchmark for LMMs), and it answers questions like “which of these three photos has the best exposure, and why?” — comparative judgment across images in free language, which single-image scoring cannot express. This is R1’s pairwise-comparison subjective protocol (the most discriminative one) reborn as an LMM capability.
• DepictQA (Zhiyuan You, Tianfan Xue, Chao Dong et al., CUHK / XPixel — a different group from Q-Future; ECCV 2024) — “Depicting Beyond Scores”: describe-and-compare image quality in natural language as a hierarchical descriptive judgment rather than a scalar. Its follow-up DepictQA-Wild scales the idea to in-the-wild data. DepictQA is the clearest statement that the description is the assessment — the score is a lossy projection of a richer linguistic verdict.

Together these mark the frontier moving from “emit a better number” to “do the quality task in language” — describe, compare, reason — which reframes what the benchmark should even measure (§Frontier).

The aesthetic arm: aesthetics and quality now share one recipe

R3 ended on VILA learning aesthetics from user comments. The MLLM era carries that forward and lands aesthetics on the same instruction-tuning recipe as quality — proof that R1’s “same machine” thesis holds on the aesthetic side too.

• Q-Align’s aesthetic arm is the simplest evidence: the identical discrete- level method, trained on AVA, gives 0.822 / 0.817 SRCC/PLCC — the top of R3’s AVA ladder — with no aesthetic-specific architecture. Quality and beauty are the same five words pointed at a different question, exactly R1’s claim.
• UNIAA (Unified Multi-modal Image Aesthetic Assessment, Kuaishou/Kling + PKU, 2024) — a unified aesthetic baseline and benchmark. UNIAA-Bench spans three levels that echo Q-Bench — aesthetic perception, description, and assessment — and UNIAA-LLaVA is the model, trained on aesthetic visual-instruction data consolidated from multiple existing IAA datasets. It is Q-Bench’s diagnosis structure ported to taste.
• AesExpert (Yipo Huang, Leida Li, Weisi Lin, Guangming Shi et al., Xidian University; ACM MM 2024) — the aesthetics instruction-tuning corpus done at scale: AesMMIT, an Aesthetic Multi-Modality Instruction Tuning dataset of 409K instructions over 21,904 images with 88K human feedbacks, and the resulting AesExpert model (LLaVA-1.5 based). It is Q-Instruct’s move — teach the model to describe — pointed at aesthetics: composition, colour harmony, mood, story, in open language.

The recipe is now visibly identical across both tracks: benchmark the MLLM’s native ability (Q-Bench / UNIAA-Bench) → instruction-tune it to describe (Q-Instruct / AesExpert / AesMMIT) → align it to rate (Q-Align, both arms). Aesthetics is no longer a separate craft; it is the quality recipe with a different training set and a richer critique.

Consolidated benchmark: one model at the top of both columns

This is the payoff table. All numbers are SRCC / PLCC, higher = better; blank = the method’s own paper does not report that column. The two tracks R2 and R3 kept apart now meet in the bottom two rows: Q-Align and OneAlign are at or above the top of the KonIQ column and the AVA column simultaneously — the literal one-model convergence.

In-domain (the convergence payoff)

Method	Era / track	KonIQ (IQA)	AVA (IAA)
MUSIQ	ViT · both	0.916 / 0.928¹	0.726 / 0.738
MANIQA	ViT · IQA (R2)	0.893²	—
CLIP-IQA+	CLIP · IQA (R2)	0.895 / 0.909	—
LIQE	CLIP · IQA (R2)	0.919 / 0.912	0.776³
VILA-R	VLP · IAA (R3)	—	0.774 / 0.774
Q-Align	MLLM	0.940 / 0.941	0.822 / 0.817
OneAlign	MLLM · unified	0.941 / 0.950	0.823 / 0.819
DeQA-Score	MLLM (2025)	0.941 / 0.953	—

The bottom block is the whole four-report arc in three rows: a single family of models tops the IQA leaderboard R2 built and the aesthetics leaderboard R3 built, and OneAlign does it with one set of weights that also scores video (LSVQ 0.886, not shown). No method above the line reports both columns competitively; every method below does.

Cross-dataset — where the generalization advantage shows (dissolving R2’s frontier)

The honest test from R2 (§Generalisation): train on one authentic set, test on another. This is where the MLLM’s external world-knowledge pays off most clearly. All rows are train on KonIQ-10k → test on the named set, SRCC:

Model	→ CLIVE	→ SPAQ	→ KADID	In-domain KonIQ (ref)
HyperIQA (CNN, R2)	0.785	—	—	0.906
CLIP-IQA+ (CLIP, R2)	0.805	—	—	0.895
Q-Align (MLLM)	0.860	0.887	0.684⁴	0.940

Read the CLIVE column top to bottom: the CNN regressor drops to 0.785 cross-dataset, the CLIP method to 0.805, and Q-Align holds 0.860 — the smallest in-domain→cross-dataset collapse in the series, and direct evidence that the pretrained prior is what generalizes. R2’s frontier problem is not solved (KADID synthetic distortions still drop it to 0.684), but it is materially dented by exactly the ingredient R2 predicted would help: language grounding on top of a massive external prior.

Footnotes. ¹MUSIQ KonIQ is 0.916 in its own paper; some re-evaluations report 0.929 (R1 used the latter) — treat sub-0.01 gaps as protocol noise. ²MANIQA’s own paper never tabulates KonIQ at paper protocol; 0.893 is the pyiqa whole-set number (optimistic, different protocol — carried from R2, flagged there). ³LIQE AVA from R1’s reference table. ⁴Q-Align KonIQ→KADID 0.684 / 0.674: KADID is synthetic, so this is a domain shift (authentic→synthetic) as much as a dataset shift. All Q-Align / OneAlign numbers are read from the Q-Align paper’s tables; DeQA-Score KonIQ from its own Table 3. Not attributed to Q-Align: FLIVE / PaQ-2-PiQ — the Q-Align paper’s tables do not report it, so despite the task’s expectation this report leaves the FLIVE column blank rather than guess. No direct MANIQA-vs- OneAlign row exists in Q-Align’s tables; its named IQA baseline is CLIP-IQA+, and the MANIQA cell above is cross-sourced from R2 for context only.

The frontier: what MLLM scoring did not solve

The convergence is a real jump, but R2/R3’s discipline was to end on the open problem, and MLLM scoring opens as many as it closes.

• Score calibration and reproducibility. A softmax over five tokens is a soft score, but it is sensitive to prompt wording, decoding temperature, and checkpoint — two runs of “the same” MLLM scorer can disagree in ways a frozen 20 MB regressor never would. DeQA-Score (You et al., CVPR 2025) attacks this directly: model the whole score distribution with a Thurstone fidelity loss rather than a point, improving calibration (KonIQ 0.941 / 0.953). Calibration is now a named subproblem, not a footnote.
• Hallucination in the explanation. The critique that makes the score auditable can also be confidently wrong — an MLLM may narrate “motion blur” that isn’t there, or rationalize a score after the fact. A fluent, plausible, incorrect reason is arguably worse than no reason, because it invites trust. The descriptive-IQA line (DepictQA, Co-Instruct) makes this failure mode first-class rather than hidden inside a scalar.
• Benchmark saturation and contamination. KonIQ in-domain is at ≈ 0.94 and AVA at ≈ 0.82 — near the noise ceiling of the human labels themselves. Worse, web-scale pretraining makes train/test contamination a live risk: if AVA or KonIQ images (or discussions of them) sat in the pretraining corpus, a high SRCC may measure memorization, not perception. This is unfalsifiable from the outside, and it undermines the very leaderboards the field optimizes.
• Compute cost. OneAlign is a multi-billion-parameter MLLM; MANIQA is ≈ 20 MB and runs on a CPU, NIQE in milliseconds with no GPU. For on-device or high- throughput scoring the MLLM is simply the wrong tool — a three-order-of-magnitude cost gap for a ~0.02–0.05 SRCC gain and an explanation you may not need.
• Is SRCC-on-AVA/KonIQ even the right target anymore? This is the deepest question the era raises. If the model can describe, compare, and reason about quality (Q-Bench, Co-Instruct, DepictQA), then reducing it to one correlation coefficient on one dataset throws away most of what it can do — and rewards contamination and overfitting. The frontier is arguably shifting from “higher SRCC” to “correct, faithful, non-hallucinated low-level description and comparison” — a benchmark like Q-Bench, not a number like AVA SRCC.

And the field’s third axis is folding in too. R1 kept generative-image quality (“how real?” — evaluating diffusion output) as a separate cousin; video quality (VQA) was a separate literature again. Both are now being absorbed by the same MLLM scorers: OneAlign already scores video (LSVQ, KoNViD); Q-Bench-Video (CVPR 2025) benchmarks LMM video-quality understanding; and Q-Eval-100K (Zhang et al., CVPR 2025 Oral — 100K text-to-image/video instances, 960K human MOS) with its Q-Eval-Score evaluator points the same machinery at generated content’s quality and prompt alignment. R1’s three orthogonal axes — degraded, beautiful, real — are converging on one linguistic scorer.

Series synthesis: the whole mental model, snapped shut

Step back across all four reports. IQA and IAA climbed the identical five-rung ladder — NSS → CNN → ViT → CLIP → MLLM — in two parallel tracks, and the two tracks were one machine the whole time. R1 asserted it as a thesis (“same MOS-prediction machine, opposite questions”); R2 and R3 walked the two tracks and showed the rungs matching (NIMA’s distribution loss served both; MUSIQ scored both; the same MAML machinery answered distortions in MetaIQA and raters in BLG-PIAA); R4 shows the tracks literally merging into one model — OneAlign, one set of weights that answers how degraded?, how beautiful?, and how good is this video? in the same five words.

The arc, one line per rung, both tracks at once:

Rung	IQA (R2)	IAA (R3)	What the rung removed
NSS / rules	BRISQUE, NIQE	Datta, Ke	— (hand-built priors)
CNN	DBCNN, HyperIQA	NIMA (EMD), MLSP	the handcrafted-feature dependency
ViT	MUSIQ, MANIQA	MUSIQ (reused)	the fixed-resolution-resize dependency
CLIP / VLP	CLIP-IQA, LIQE	VILA	the labelled-MOS dependency
MLLM	Q-Align	Q-Align (AVA arm)	the bespoke-model-per-task dependency

Each rung removed one dependency of the last; the final rung removed the biggest one — the need for a separate model at all. R1’s dotted line is now a solid checkpoint.

But the crown of the mental model is knowing what to reach for. “Use the MLLM” is wrong more often than it is right. The decision guide, by constraint:

• Need ≥ 0.9 SRCC and interpretability, compute available (server / batch): Q-Align / OneAlign (or DeQA-Score when calibration matters most). You get the top of both leaderboards, cross-dataset robustness, and a natural-language reason. This is the right default only when you can afford billions of parameters and want the explanation.
• Need tiny / fast / on-device, or high throughput: MANIQA or HyperIQA (≈ 0.90 KonIQ, a few MB–tens of MB, GPU-optional), or NIQE when you need zero training, zero labels, and millisecond CPU scoring and can accept ≈ 0.66 on authentic photos. The MLLM’s ~0.02–0.05 SRCC edge does not justify a 1000× cost here.
• Need a perceptual training loss (super-resolution, diffusion, codecs): LPIPS or DISTS (R2’s parallel FR branch) — differentiable, reference- based, and the field’s standard. MLLM scorers are not differentiable losses; do not reach for them here.
• Need aesthetics specifically: the same tiering — NIMA (distribution, tiny) or MLSP / VILA (native-res / comment-pretrained, mid-size) for cheap scoring; Q-Align or AesExpert / UNIAA when you want the score plus an open-language critique of composition and mood.
• Need to compare images, or a reason, not a number: Co-Instruct (multi-image comparison) or DepictQA (descriptive judgment) — the tasks a regressor structurally cannot do.

That table and that guide are the deliverable of the whole series: not “MLLMs won”, but a map of a single problem — predict a human opinion of an image — with a five-era ladder, two tracks that are one machine, and a clear-eyed sense of which rung to stand on for which job. R1 drew the map; R2 and R3 walked the two paths; R4 is where they meet, and where the practitioner picks a tool with the whole picture in view.

Sources

The Q-family (Q-Future · NTU / SJTU, Wu et al.). Q-Bench (Wu, Zhang, Zhang, Chen, Liao, Wang, Li, Sun, Yan, Zhai, Lin; ICLR 2024 Spotlight, arXiv:2309.14181) (code) · Q-Instruct (Wu, Zhang, Zhang, Chen, Liao, Wang, Xu, Li, Hou, Zhai, Xue, Sun, Yan, Lin; CVPR 2024, arXiv:2311.06783) (code) · Q-Align / OneAlign (Wu, Zhang, Zhang, Chen, Liao, Li, Gao, Wang, Zhang, Sun, Yan, Min, Zhai, Lin; ICML 2024, arXiv:2312.17090) (code) · Co-Instruct (Wu, Zhu, Zhang, Zhang, Chen, Liao, Li, Wang, Sun, Yan, Liu, Zhai, Wang, Lin; ECCV 2024 Oral, arXiv:2402.16641) (code).

Descriptive / comparative IQA (other groups). DepictQA — “Depicting Beyond Scores” (You, Li, Gu, Yin, Xue, Dong; ECCV 2024, arXiv:2312.08962) (code), follow-up DepictQA-Wild (arXiv:2405.18842) · Compare2Score (Zhu, Wu, Li, Zhang, Chen, Zhu, Fang, Zhai, Lin, Wang; NeurIPS 2024, arXiv:2405.19298).

Aesthetic arm. UNIAA (Zhou, Wang, Lin, Su, Chen, Tao, Zheng, Yuan, Wan, Zhang; arXiv:2404.09619) (code) · AesExpert / AesMMIT (Huang, Sheng, Yang, Yuan, Duan, Chen, Li, Lin, Shi; ACM MM 2024, arXiv:2404.09624) (code).

Calibration / reasoning / generative-and-video frontier. DeQA-Score (You, Cai, Gu, Xue, Dong; CVPR 2025, arXiv:2501.11561) (code) · Q-Insight (Li, Zhang, Zhao, Zhang, Li, Zhang, Zhang; 2025, arXiv:2503.22679) · Q-Eval-100K / Q-Eval-Score (Zhang, Kou, Wang, Li, Sun, … Liu, Zhai; CVPR 2025 Oral, arXiv:2503.02357) (code) · Q-Bench-Video (Q-Future; CVPR 2025, arXiv:2409.20063) (code).

Prior reports in this series. R1 — Foundations · R2 — IQA methods (NSS to CLIP) · R3 — IAA methods (NIMA to VILA). Concept pages: Multimodal-LLM Visual Scoring, Image Quality Assessment (IQA), Image Aesthetic Assessment (IAA), IQA / IAA Datasets, IQA / IAA Evaluation Metrics.

Benchmark numbers. Q-Align / OneAlign SRCC/PLCC (KonIQ, SPAQ, KADID, AVA, LSVQ, KoNViD; cross-dataset KonIQ→CLIVE/SPAQ/KADID) and the levels-as-tokens equation and ablations (Tables 1, 4, 11) are read from the Q-Align paper. DeQA-Score KonIQ 0.941/0.953 from its Table 3. R2/R3 comparison rows (MUSIQ, MANIQA, CLIP-IQA+, LIQE, VILA-R) carry their flags from those reports.

Flagged as not fully verified against a primary source (stated, not asserted as fact): Compare2Score and Q-Insight author lists were verified from a single search-result source, not a direct arXiv author-block fetch — names may be incomplete or mis-ordered. Q-Insight carries the “Q-“ name but is a different group (Jian Zhang et al., not Q-Future) — do not attribute it to Haoning Wu’s cluster. FLIVE / PaQ-2-PiQ is not reported in Q-Align’s tables, so no OneAlign FLIVE number is stated here despite the series’ interest in that column. MANIQA does not appear as a row in Q-Align’s IQA tables (its named IQA baseline is CLIP-IQA+); the MANIQA KonIQ cell is a cross-sourced pyiqa whole-set number from R2, on a different, optimistic protocol. UNIAA has no confirmed peer-reviewed venue (arXiv-only as verified). Corresponding-author designations (Weisi Lin on the three Q-papers) are inferred from consistent last-author placement, not from a marked asterisk in the fetched abstracts. Weixia Zhang is an author on Q-Align only, not on Q-Bench or Q-Instruct — the “canonical Q-cluster” author list is over-broad on that name.