Aarna004/vlm-distortion-sensitivity

Beyond Correlation: Measuring the Distortion Sensitivity of Vision-Language Models for Image Quality Assessment

This dataset contains 28 new pristine images and 28,000 distorted versions: ten distortion types, each applied at 100 intensity levels (k = 1 … 100, higher k = stronger distortion). It was built to measure how the quality scores of image quality assessment (IQA) models change as a distortion gradually intensifies, beyond the correlation scores reported on standard benchmarks.

Pristine reference and its ten distorted versions at k = 50 for an indoor image (top) and an outdoor image (bottom).

Dataset structure

reference_images/
  QP_IMG_<n>.png              28 pristine images, 1024 x 768, lossless PNG
  metadata.csv                file_name, scene_category, scene_subcategory, width, height
distorted_images/
  D_IMG_<n>/<CODE>_k_<k>.webp distorted versions of QP_IMG_<n>, lossless WebP (10 codes x 100 levels per image)
  metadata.csv                file_name, reference, distortion, k, parameter, parameter_value

All images are stored losslessly; the only compression artifacts are those of the JPEG and JPEG 2000 distortions.

from datasets import load_dataset
distorted = load_dataset("Aarna004/vlm-distortion-sensitivity", "distorted", split="test")
reference = load_dataset("Aarna004/vlm-distortion-sensitivity", "reference", split="test")

Pristine images

28 photographs (11 indoor, 17 outdoor), resized and center-cropped to 1024 x 768 pixels (the resolution of KonIQ-10k).

Category	Images	Subcategories
Indoor	11	atrium (3), church interior (2), casino, corridor, dining table, museum, office, staircase
Outdoor	17	building exterior (2), mountain (2), covered station (2), church exterior, desert landscape, garden at night, lakefront city, lakeshore, park, pond, resort, riverside at night, street, street at night

Distortions

Each distortion is controlled by a level k = 1 … 100. The distortion parameter moves from p(1) to p(100) along a linear, logarithmic, or square-root interpolation; the end points were chosen empirically to cover the range of noticeable variation, from a barely visible distortion at k = 1 to a severe distortion at k = 100.

Code	Distortion	Parameter	Mapping	p(1)	p(100)
GB	Gaussian blur	sigma (px)	linear	0.3	10
MB	Motion blur (30°)	length (px)	linear	2	60
GN	Gaussian noise	sigma (0-255)	linear	1	60
IN	Impulse noise	density (fraction of pixels)	log	0.005	0.25
JPEG	JPEG compression	quality (1-100)	sqrt (gamma 0.5)	75	3
J2K	JPEG 2000 compression	ratio (compression ratio)	log	30	600
CTR	Contrast reduction	alpha (remaining contrast)	linear	0.98	0.1
OVE	Overexposure	gain (exposure multiplier)	linear	1.03	2
PIX	Pixelation	block (px)	linear	1.2	16
SHP	Oversharpening (unsharp mask)	amount (unsharp-mask weight)	linear	0.2	10

QP_IMG_18 under each distortion at k = 1, 20, 40, 60, 80, 100.

Results

Seven vision-language models (VLMs) were evaluated with their released weights, preprocessing, and prompts, without training or fine-tuning: the CLIP-based LIQE, QualiCLIP, and FGResQ; the generative VLMs Q-Align and Q-ReAlign; and the reasoning VLMs Q-Insight and VisualQuality-R1. The two reasoning VLMs generate their scores as text with greedy decoding and are run without the reasoning chain. Each score is mapped to [0, 1] with the model's nominal output range [a, b] and divided by the score of the pristine image, giving the relative score r(k): the fraction of the pristine-image score that remains at level k (r(0) = 1). VisualQuality-R1 occasionally outputs scores below its nominal minimum of 1; to preserve their order, these are mapped linearly to [0, 0.1) and scores in [1, 5] to [0.1, 1].

Model	Type	Nominal range [a, b]
LIQE	CLIP-based	[1, 5]
QualiCLIP	CLIP-based	[0, 1]
FGResQ	CLIP-based	[0, 1]
Q-Align	Generative	[1, 5]
Q-ReAlign	Generative	[0, 1]
Q-Insight	Reasoning	[1, 5]
VisualQuality-R1	Reasoning	[1, 5]

Each response curve r(k), k = 0 … 100, is summarized by:

• S_lin, linear sensitivity: −100 × the slope of a least-squares line, i.e. the fraction of the pristine-image score lost over the full range (higher = more sensitive).
• R²_lin: coefficient of determination of that line (low = the decline is not spread evenly over the range).
• k50, transition level: k at which a fitted decreasing logistic curve has fallen halfway between its plateaus ("<0": halfway point before the first level; ">100": after the last level).
• S_peak, peak sensitivity: steepest decline of the fitted logistic within k = 0 … 100, per 100 levels.

A logistic fit is valid when its two plateaus differ by at least 0.1 and its coefficient of determination is at least 0.8; k50 and S_peak are reported for valid fits only ("n/a" when no image gives a valid fit). Values are medians over the 28 pristine images. Since k maps to a different parameter for each distortion type, measures should be compared between models within a distortion type (column), not between distortion types.

Median response curves over the 28 pristine images for each distortion type. Line style indicates the model type: solid for CLIP-based, dashed for generative, and dotted for reasoning VLMs.

Per-model summary

Averages over the ten distortion types. Rank: mean rank of S_lin within each distortion type (1 = most sensitive).

Model	Type	S_lin	Rank	R²_lin	S_peak
LIQE	CLIP-based	0.69	2.9	0.70	4.37
QualiCLIP	CLIP-based	0.45	5.1	0.69	4.23
FGResQ	CLIP-based	0.30	6.2	0.91	0.66
Q-Align	Generative	0.70	2.7	0.87	1.77
Q-ReAlign	Generative	0.54	3.9	0.79	2.06
Q-Insight	Reasoning	0.54	4.4	0.86	1.33
VisualQuality-R1	Reasoning	0.67	2.8	0.83	2.05

Linear sensitivity S_lin — median (inter-quartile range)

Model	Type	GB	MB	GN	IN	JPEG	J2K	CTR	OVE	PIX	SHP
LIQE	CLIP-based	0.53 (0.06)	0.25 (0.06)	0.95 (0.13)	0.78 (0.18)	1.22 (0.22)	0.42 (0.26)	1.02 (0.10)	0.62 (0.45)	0.30 (0.11)	0.82 (0.12)
QualiCLIP	CLIP-based	0.51 (0.08)	0.17 (0.06)	0.59 (0.14)	0.54 (0.15)	0.87 (0.06)	0.34 (0.20)	0.59 (0.12)	0.26 (0.12)	0.24 (0.06)	0.43 (0.17)
FGResQ	CLIP-based	0.55 (0.07)	0.28 (0.04)	0.54 (0.12)	0.24 (0.05)	0.16 (0.06)	0.39 (0.10)	0.24 (0.05)	0.06 (0.05)	0.28 (0.05)	0.28 (0.05)
Q-Align	Generative	0.80 (0.06)	0.55 (0.07)	0.87 (0.09)	0.61 (0.13)	0.66 (0.15)	0.94 (0.13)	0.80 (0.08)	0.30 (0.30)	0.93 (0.08)	0.52 (0.22)
Q-ReAlign	Generative	0.55 (0.08)	0.40 (0.11)	0.71 (0.08)	0.76 (0.18)	0.64 (0.09)	0.70 (0.15)	0.39 (0.11)	0.04 (0.11)	0.56 (0.12)	0.62 (0.70)
Q-Insight	Reasoning	0.84 (0.05)	0.61 (0.05)	0.62 (0.08)	0.51 (0.09)	0.49 (0.06)	0.68 (0.22)	0.38 (0.07)	0.12 (0.14)	0.97 (0.11)	0.22 (0.11)
VisualQuality-R1	Reasoning	0.90 (0.09)	0.67 (0.07)	0.67 (0.11)	0.79 (0.21)	0.46 (0.10)	0.94 (0.14)	0.58 (0.15)	0.23 (0.25)	1.11 (0.07)	0.38 (0.21)

Transition level k50 — median

Model	Type	GB	MB	GN	IN	JPEG	J2K	CTR	OVE	PIX	SHP
LIQE	CLIP-based	2	<0	2	91	78	<0	27	52	3	<0
QualiCLIP	CLIP-based	5	<0	11	50	70	<0	>100	74	<0	<0
FGResQ	CLIP-based	<0	<0	<0	58	34	82	48	71	25	<0
Q-Align	Generative	<0	<0	33	76	94	49	78	59	18	21
Q-ReAlign	Generative	4	<0	14	68	89	34	>100	94	16	58
Q-Insight	Reasoning	5	<0	40	>100	94	71	85	70	37	18
VisualQuality-R1	Reasoning	1	2	68	91	>100	59	>100	77	30	41

Median over the 28 pristine images of (a) the linear sensitivity S_lin and (b) the transition level k50, for each model (rows) and distortion type (columns). In (b), the color scale is clipped to [0, 100]; "<0" and ">100" mark transitions outside the evaluated range.

Linearity R²_lin — median

Model	Type	GB	MB	GN	IN	JPEG	J2K	CTR	OVE	PIX	SHP
LIQE	CLIP-based	0.44	0.24	0.85	0.93	0.94	0.59	0.97	0.92	0.26	0.83
QualiCLIP	CLIP-based	0.52	0.18	0.72	0.97	0.87	0.66	0.89	0.96	0.26	0.91
FGResQ	CLIP-based	0.91	0.76	0.90	0.97	0.90	0.97	0.98	0.95	0.80	0.92
Q-Align	Generative	0.82	0.71	0.94	0.96	0.71	0.98	0.96	0.97	0.71	0.97
Q-ReAlign	Generative	0.62	0.51	0.95	0.95	0.69	0.95	0.83	0.76	0.65	0.96
Q-Insight	Reasoning	0.91	0.69	0.98	0.87	0.60	0.96	0.92	0.84	0.94	0.93
VisualQuality-R1	Reasoning	0.82	0.64	0.97	0.84	0.54	0.94	0.92	0.85	0.82	0.92

Peak sensitivity S_peak — median

Model	Type	GB	MB	GN	IN	JPEG	J2K	CTR	OVE	PIX	SHP
LIQE	CLIP-based	6.99	13.26	2.53	1.37	1.96	2.85	1.42	1.09	9.79	2.46
QualiCLIP	CLIP-based	4.64	17.88	2.16	0.73	1.99	2.15	1.62	0.39	9.73	0.98
FGResQ	CLIP-based	1.27	1.16	1.10	0.32	0.26	0.56	0.32	0.11	0.76	0.73
Q-Align	Generative	2.68	3.39	1.33	0.83	2.69	1.20	1.19	0.47	3.25	0.73
Q-ReAlign	Generative	3.59	5.19	1.14	1.06	2.94	1.13	1.49	0.31	2.71	1.02
Q-Insight	Reasoning	1.51	3.16	0.72	1.24	2.78	1.01	0.60	0.34	1.60	0.34
VisualQuality-R1	Reasoning	2.41	4.06	0.87	1.89	4.35	1.52	0.99	0.61	3.22	0.56

Median over the 28 pristine images of (a) the coefficient of determination R²_lin of the linear fit and (b) the peak sensitivity S_peak, for each model (rows) and distortion type (columns).

Response curves per pristine image

Response curves r(k) of each pristine image, without the median over images. Click an image to expand it.

QP_IMG_1 — Outdoor, building exterior

QP_IMG_2 — Outdoor, pond

QP_IMG_3 — Indoor, dining table

QP_IMG_4 — Indoor, church interior

QP_IMG_5 — Indoor, office

QP_IMG_6 — Outdoor, street at night

QP_IMG_7 — Outdoor, resort

QP_IMG_8 — Outdoor, desert landscape

QP_IMG_9 — Indoor, museum

QP_IMG_10 — Indoor, casino

QP_IMG_11 — Outdoor, garden at night

QP_IMG_12 — Indoor, atrium

QP_IMG_13 — Indoor, atrium

QP_IMG_14 — Indoor, staircase

QP_IMG_15 — Indoor, atrium

QP_IMG_16 — Indoor, corridor

QP_IMG_17 — Outdoor, riverside at night

QP_IMG_18 — Outdoor, mountain

QP_IMG_19 — Outdoor, covered station

QP_IMG_20 — Outdoor, covered station

QP_IMG_21 — Outdoor, mountain

QP_IMG_22 — Indoor, church interior

QP_IMG_23 — Outdoor, church exterior

QP_IMG_24 — Outdoor, lakeshore

QP_IMG_25 — Outdoor, lakefront city

QP_IMG_26 — Outdoor, street

QP_IMG_27 — Outdoor, park

QP_IMG_28 — Outdoor, building exterior

Conclusions

• Most and least sensitive models. Q-Align, VisualQuality-R1, and LIQE are the most sensitive (mean rank 2.7 to 2.9, average S_lin 0.67 to 0.70); Q-ReAlign and Q-Insight are intermediate; QualiCLIP and FGResQ are the least sensitive.
• Model type does not determine sensitivity. The CLIP-based models include both LIQE and FGResQ, and the reasoning VLMs rank second (VisualQuality-R1) and fifth (Q-Insight).
• Shape of the response. LIQE and QualiCLIP react abruptly to weak distortions and then saturate, combining the lowest R²_lin with the highest S_peak; Q-Align and FGResQ decline evenly over the range.
• Location of the decline depends on the distortion type. Blur lowers all scores within the first few levels (k50 < 6 for every model), whereas JPEG compression and impulse noise affect them mainly at high levels (k50 ≥ 50 for every model except FGResQ on JPEG). For Gaussian noise, JPEG 2000, pixelation, and oversharpening, the location depends on the model.
• Correlation is not enough. Correlation-based benchmark scores do not characterize how a model responds to gradual quality changes: a model that saturates early cannot separate moderate from severe distortion, and a model with a late transition cannot separate pristine images from mildly distorted ones.
• Choosing a model. Applications that compare images of similar quality, such as parameter tuning in restoration or compression, require a model whose response covers the relevant intensity range; the sensitivity profiles support this choice.

License

The images are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license: they may be shared and adapted for non-commercial purposes with appropriate credit.

Citation

@misc{nema2026beyondcorrelation,
  title  = {Beyond Correlation: Measuring the Distortion Sensitivity of Vision-Language Models for Image Quality Assessment},
  author = {Nema, Arpita and Lin, Weisi},
  year   = {2026}
}