Is VDP the same as PSNR?

No. PSNR is based on numerical error magnitude, while a VDP incorporates a model of perceptual visibility.

Does every VDP handle video or color in the same way?

No. Supported content and perceptual dimensions vary by model. Choose and document the implementation appropriate to the task.

Does a visible difference always mean worse quality?

No. The test image may contain an intentional or beneficial change. Visibility alone does not assess preference or purpose.

AI Glossary / Computer vision and multimodal AI / Intermediate

Visible differences predictor (VDP)

Definition

A visible differences predictor (VDP) is a perceptual image-comparison model that estimates whether people would notice differences between a reference image and a test image under specified conditions. A typical output is a spatial map of predicted difference visibility, rather than a simple count of changed pixels.

What it means

Two images can differ numerically without looking obviously different. Human sensitivity depends on factors such as brightness, detail scale and surrounding texture. A VDP attempts to account for such factors when predicting where a change becomes perceptible.

The name is associated with Scott Daly's Visible Differences Predictor and subsequent visual-difference models, including HDR-VDP variants. These are related methods, not one interchangeable implementation. A report should identify the particular model and version.

Difference visibility is also distinct from overall quality or preference. A visible change might be an improvement, an unwanted artifact or an intentional design change. A prediction that a difference is noticeable does not determine which interpretation applies.

How it works

A model receives aligned reference and test images together with the inputs needed to describe their presentation. Depending on the implementation, these can include image encoding, absolute luminance and angular resolution in pixels per visual degree.

Its visual-system model processes the images to represent sensitivity to luminance, spatial structure and masking by other content. The comparison is then mapped to a prediction of visibility. Some model families also provide separate quality or contrast-distortion outputs.

The meanings of these outputs must be kept distinct. A local detection probability, a pooled detection value and a quality score do not share one universal scale or aggregation rule.

The prediction is conditional on the model and viewing setup. Incorrect luminance interpretation, misalignment or an unsuitable display assumption can make the output answer a different question from the one the team intended.

Three practical examples

An image-compression experiment examines where a candidate encoding is predicted to introduce visible changes relative to the reference.
A rendering team compares a faster approximation with a high-quality reference and reviews the predicted difference regions.
A display-evaluation study tests whether measured image differences are likely to be visible under the stated viewing conditions.

Think of it this way

Think of a map that highlights where a reader is likely to notice changes between two printed pages. It directs attention to potentially visible differences without deciding whether the edits improve the writing.

The analogy is limited because a VDP models visual appearance, not meaning. A tiny changed digit can matter greatly even if its visibility is low, while a conspicuous decorative change may be harmless.

Worked example

Interpreting a probability map

Imagine a test produces four illustrative local visibility predictions: 0.10, 0.20, 0.80 and 0.90. With a chosen review threshold of 0.75, the last two locations are flagged for inspection. The threshold is a workflow choice, not a universal definition of acceptable quality.

The arithmetic mean is 0.50. That does not automatically mean there is a 50% probability that a viewer will notice any difference in the complete image. The predictions are spatially related, and a model's pooling rule must be used for any supported global interpretation.

Similarly, two flagged locations out of four do not establish that half of viewers dislike the result. Visibility and preference are different questions, and these locations do not represent a sampled viewer population.

Now repeat the comparison with a different viewing-distance setting. A suitable visibility model may change its predictions because the same pixel pattern subtends a different visual angle. The image files have not changed, but the presentation question has.

These invented values illustrate interpretation; no VDP implementation was run.

Illustrative case study

Illustrative case study - fictional

A fictional design platform compares two export settings for detailed illustrations. One produces a smaller file, and a VDP map highlights differences around fine linework. The team inspects those areas at the intended display size rather than judging only a magnified crop.

It then discovers that its first run used an incorrect interpretation of the image's luminance encoding. The comparison is repeated with the intended settings, and a small human review checks whether the remaining changes affect the illustrations' use.

This illustrates why model inputs matter. Perceptual predictions do not replace user testing or guarantee preference.

Common misconceptions and mistakes

A VDP is generally a reference-based comparison, not an assessment of authenticity or factual correctness.
Predicted invisibility is conditional, not a promise that no person under any conditions can notice the change.
A probability map is not automatically an overall image-quality score.
Averaging local probabilities does not automatically produce a valid global detection probability.
Different versions can have different inputs, calibration and outputs. Their numerical results should not be compared as if they were identical scales.

Business use

Businesses can use visibility prediction to focus review and compare processing choices under controlled conditions. It can help explain why equal pixel errors do not always have equal perceptual effects.

Record the reference, test image, model version, presentation assumptions and output being interpreted. Inspect important content separately, especially text or small features whose meaning matters. When selecting a threshold, validate it against the intended audience and viewing conditions. Treat the model as evidence within an evaluation process, not as a universal approval stamp.

Check your understanding

A local visibility map averages to 0.5. Can you report that half of viewers will dislike the image?

Answer: No. The average is not a preference measure and is not automatically a valid pooled detection probability.

Related terms

HDR-VDP
Structural similarity (SSIM)
perceptual image quality
contrast sensitivity
visual masking
PSNR
structural similarity

Learn next

Compare a reference with versions containing blur, noise and a small text change. Document the viewing settings and inspect each model output separately. Compare predictions with observations at the intended display size before choosing a review threshold.

Sources

Sources: Mantiuk, Myszkowski and Seidel, Visible Difference Predicator for High Dynamic Range Images: https://www.cl.cam.ac.uk/~rkm38/pdfs/mantiuk04VDPForHDR.pdf ; HDR-VDP project FAQ: https://hdrvdp.sourceforge.net/wiki/index.php/Frequently_Asked_Questions ; Mantiuk, Hammou and Hanji, HDR-VDP-3 overview: https://arxiv.org/abs/2304.13625 . Relevant original-paper text, output explanations and research abstract were reviewed. No model was run for the illustrative values.

Previous: Visibility mask Next: Visual analytics