AI Glossary / Computer vision and multimodal AI / Advanced

HDR-VDP

Definition

HDR-VDP is a family of visual-difference prediction methods designed to estimate how people perceive differences between images across a broad luminance range. It models aspects of human vision rather than treating every numerical pixel difference as equally visible or equally important.

What it means

Two images can have the same average pixel error yet look very different to a viewer. A small change in a smooth dark region may be noticeable, while a larger change can be hidden by surrounding texture. HDR-VDP uses a perceptual model to account for such effects. The name expands to High Dynamic Range Visual Difference Predictor. Versions provide different outputs and capabilities; HDR-VDP-3, for example, addresses visibility, quality and contrast-distortion tasks. The exact version, output and viewing assumptions must be stated before interpreting a result. A predicted perceptual difference is not the same as a judgment about whether depicted content is true or useful.

How it works

The method compares a reference image with a test image under specified display or luminance and viewing conditions. Its visual model accounts for factors such as sensitivity at different spatial frequencies, adaptation and the masking of differences by image content. The processing produces task-dependent maps or summary measures, according to the chosen version and mode.
Inputs need the correct interpretation. Scene values, display-encoded pixels and absolute luminance values cannot be substituted indiscriminately. Viewing geometry determines how pixels map to visual angle, which can affect predicted visibility. A full-reference comparison also assumes the images correspond appropriately. Misalignment can produce differences that reflect registration error rather than the intended compression or rendering distortion.

Three practical examples

• An imaging researcher compares a compressed HDR image with its reference and studies where visible artifacts are predicted.
• A rendering team evaluates whether an approximation creates perceptible changes under a defined display and viewing setup.
• A quality-assessment study compares metric predictions with human judgments, checking whether the selected version and task match the experiment.

Think of it this way

Think of an inspector who considers not only how much two measurements differ but also whether the difference is likely to stand out in its surroundings. The analogy is limited because HDR-VDP is a mathematical model fitted and evaluated for particular perceptual tasks, not a human observer who understands every image or application context.

Worked example

Suppose two image-processing methods introduce identical numerical errors into different regions. Method A changes a smooth patch, while method B changes a highly textured patch. A simple mean-squared-error calculation may treat the equal errors alike. A perceptual comparison can predict different visibility because texture may mask some changes and viewing conditions affect sensitivity.
To evaluate the methods, a researcher uses the same reference, specifies the display and viewing parameters, runs the same HDR-VDP version and output mode, and compares the resulting maps with actual crops. This is a worked evaluation design, not an invented numerical HDR-VDP result. Real score values require running the implementation with fully specified inputs.

Illustrative case study

Fictional example: Alder Display compares two compression settings for a demonstration sequence. Its first metric run uses normalized image values as though they were calibrated luminance, producing an inappropriate interpretation. The team corrects the input handling, records the viewing setup and reviews predicted artifact regions with observers. It keeps the quality score separate from visibility maps. The case illustrates configuration and validation, not a claim that one perceptual metric can replace all viewing tests.

Common misconceptions and mistakes

HDR-VDP is not a universal truth about every viewer. Different versions and output modes should not be compared as though their numbers had identical meaning. A high quality prediction does not establish semantic correctness or absence of factual image alterations. Incorrect luminance scaling can invalidate the intended interpretation. Viewing distance and display conditions are not optional context when the model depends on them. Finally, a scalar summary can conceal a locally important difference, so spatial maps and application-specific review may still be needed.

Business use

Businesses evaluating image pipelines should define the perceptual task before selecting a metric. Record the reference, image encoding, luminance assumptions, viewing geometry, software version and output mode. Use representative content and confirm that predicted rankings align with relevant human judgments. A metric can help prioritize engineering work, but brand-critical text, product details or task-specific cues may require separate checks. Avoid publishing a score without enough context to reproduce it or understand whether higher or lower values indicate improvement for that particular output.

People also ask

Does HDR-VDP work only on HDR content?

The family is designed for broad luminance conditions, and version-specific methods can also address regular content when inputs are configured correctly.

Is a visibility map the same as a quality score?

No. Predicting whether differences are noticeable and predicting overall quality are related but distinct tasks.

Can I enter arbitrary normalized pixels without calibration?

Only according to the implementation's documented input mode. Incorrect interpretation of brightness or luminance can lead to misleading predictions.

Check your understanding

Two teams report HDR-VDP numbers using different versions, output modes and viewing assumptions. Are the values directly comparable?

Answer: Not without reconciling those conditions. The reported quantity and its interpretation may differ, so the evaluation setup must be specified.

Related terms

• HDR (high dynamic range)
• Full-reference metric
• image quality assessment
• perceptual metric
• HDR
• visual masking
• luminance
• viewing conditions

Learn next

Learn pixel-error metrics and their perceptual limitations, then distinguish visibility prediction from quality prediction. Read the chosen HDR-VDP version's input requirements before running controlled comparisons and interpreting its maps or scores.

More in Computer vision and multimodal AI

• Hausdorff distance - Hausdorff distance measures the greatest nearest-neighbor separation between two sets of points.
• Hays - Efros scene completion - Hays - Efros scene completion is a research approach that fills missing image regions by retrieving compatible content from a very large photograph collection.
• HDR merging - HDR merging combines multiple measurements of the same scene to produce an image representation with useful information across a wider brightness range.
• Healing brush - A healing brush is an image-retouching tool that replaces unwanted local detail while adapting the replacement to the surrounding appearance.
• Hierarchical hidden Markov model - A hierarchical hidden Markov model, or HHMM, is a probabilistic sequence model whose hidden states are organized into nested levels.
• High-pass filter - A high-pass filter attenuates slowly varying components of a signal while retaining or emphasizing more rapidly varying components.

Sources

Mantiuk, Hammou and Hanji, HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions: https://arxiv.org/abs/2304.13625 HDR-VDP project FAQ, including version-specific cautions: https://hdrvdp.sourceforge.net/wiki/index.php/Frequently_Asked_Questions Primary paper abstract and official project excerpts consulted; full implementation-specific and editorial review remain pending. No metric was executed and no experimental scores are claimed.

Previous: HDR merging Next: Head