Image similarity and product accuracy: where a quality score stops

A persuasive product image can contain an incorrect connector, an extra seam or a changed letter. The overall impression may still be convincing. For a marketing team, this is a frustrating combination: the image is attractive, yet it cannot be used.

Similarity metrics can formalise part of the comparison between images. The buyer still needs to know what decision a number supports. Evaluating perceived quality, checking a product model and confirming approved packaging artwork are different tasks.

Research question and scope

This review asks how evidence about perceptual metrics can inform AI/CGI purchasing without allowing an overall score to replace product acceptance.

Three primary research publications were selected: LPIPS, DISTS and A-DISTS. Selection required a described image-comparison approach, an accessible author publication and a meaningful distinction in the treatment of features, structure or texture. Source pages were checked on 27 September 2026. This is a comparison of stated purposes and principles, not a reproduction of the authors' experiments.

The scope is an individual image and its commercially important regions. Motion coherence, identity across frames and benchmarking video generators require separate procedures. The acceptance matrix below is an original purchasing framework.

What the primary papers establish

The LPIPS paper examines the relationship between deep features and human judgements of perceptual similarity. Its findings concern the authors' evaluation data and method comparisons. The first version was submitted on 11 January 2018; the checked revision is dated 10 April 2018. Zhang and colleagues, LPIPS.

DISTS combines structure and texture assessment. Its authors deliberately include tolerance to texture resampling and describe relative insensitivity to some geometric transformations. The first version appeared on 16 April 2020; the checked revision is dated 16 December 2020. Ding and colleagues, DISTS.

A-DISTS addresses the limitations of a global assessment through adaptation to locally structured and textured regions.

Submitted on 16 October 2021, it illustrates why an overall measure and the content of individual regions need separate consideration. Ding and colleagues, A-DISTS.

These findings do not provide a universal acceptance threshold for an advertising image of a particular product. Naming a metric in a brief cannot create such a threshold.

Define the meaning of an error first

A small change in a decorative stone background may be acceptable. A similar change in a proprietary textile pattern may alter the identity of the item. Both changes involve texture, but their commercial consequences differ.

Describe errors through the role of the region. Product text requires correct content. The body silhouette requires an approved shape. A reflection may change with the lighting. A defocused background can usually allow more creative freedom. The brief should contain these distinctions before anyone discusses a score.

The properties of the selected approaches support a purchasing hypothesis: a measure useful for overall perception may have a different sensitivity from that required for a critical product detail. This review has not measured the size of that difference. The practical response is to identify critical features and check them explicitly.

A product acceptance matrix

Name and marking. Decision: Correct content, order and position. Evidence for review: Approved artwork and enlarged region. Useful rejection note: Identify the wrong character or position.

Outline and proportions. Decision: Correct visible shape. Evidence for review: Agreed angle, drawing or model. Useful rejection note: Name the altered component.

Functional feature. Decision: Correct presence and location. Evidence for review: Product specification. Useful rejection note: Identify the missing or displaced feature.

Proprietary pattern. Decision: Correct recognisable elements. Evidence for review: Approved pattern reference. Useful rejection note: Mark the replaced detail.

Creative environment. Decision: Fit with art direction. Evidence for review: Approved style frame. Useful rejection note: Describe the specific mismatch.

The final column protects both parties. A comment that the product does not look right is difficult to act on. A comment that the approved model has two ventilation slots while the image has three identifies a concrete correction.

The matrix also clarifies who should review each row. A product specialist can approve functional features; the brand team can approve artwork; a creative lead can approve the environment. Consolidating their decisions is more useful than averaging unrelated preferences.

A score needs a documented comparison procedure

If a supplier proposes automatic assessment, request the method name, compared files and image preparation conditions. The comparison region, resolution, selected reference and interpretation of the result should be explicit. A number without these details is difficult to reproduce.

Define what constitutes the same comparison task. A new viewing angle or lighting treatment can introduce intended differences. The result should not automatically be treated as a defective copy of a control image. First decide which differences are allowed; then choose how to observe the remaining requirements.

Metrics may be useful for preliminary sorting or identifying changes between approved stages. Product acceptance can still use the feature map. In that arrangement, automation directs attention while the purchasing decision remains connected to the item being represented.

A supplier may also decide that manual inspection is proportionate for a small set of critical images. The presence of an automated number is not itself a reason to prefer one proposal. What matters is whether the review method addresses the failure that would prevent publication.

Hypothetical example: an industrial sensor render

Suppose a team prepares several advertising images of an industrial sensor. One has the strongest composition; another depicts the housing more accurately. The most attractive option has a displaced connector and an abbreviated model designation.

The acceptance map sets mandatory conditions: the housing matches the selected variant, the connector appears on the correct side and the complete model designation is reproduced. Background treatment and reflections remain creative choices. The image with the wrong connector fails the product requirement regardless of its overall appeal.

A marketing manager does not need to understand a neural network to make this decision. The project needs an owner of product information and an approved source. If that source is missing, resolve the data uncertainty first; otherwise, the team may polish an image of an unspecified product version.

The review record should retain the approved reference and the final decision together. A later reviewer can then distinguish a deliberate representational simplification from an unnoticed defect.

Separate correction causes

Distinguish a representation error, a source-data change and a new creative request. Distorted approved marking is a correction against the agreed source. A product rename after approval changes the source requirement. A preference for a different lighting treatment changes the creative direction.

This distinction makes scope easier to discuss than an argument about whether high quality was promised. For an image series, the issue log should name the common cause and affected versions. If the underlying model contains an incorrect feature, correct that source and then review related outputs.

This matters in CGI product video, where an asset may appear repeatedly. In a hybrid workflow, also identify where the product remains a controlled element and where it changes with the generated environment. That decision affects what needs inspection.

Request the right evidence before commissioning the series

Choose a visual test that exposes the important feature: a readable label, a difficult outline or a distinctive pattern. An attractive wide shot may conceal the main production risk.

The review package can include the complete image and close views of critical regions. Agree the viewing scale in advance. A detail cannot simultaneously remain inconspicuous in the composition and be perfectly readable under arbitrary enlargement unless that requirement has been designed into the shot.

After the test, record what the selected approach preserves, what needs a separate controlled layer and which scenes should be reconsidered. This turns acceptance into a production decision. It can change the proposed treatment before the same unresolved problem spreads across a larger package.

Limits and purchasing conclusion

This review does not establish that one metric is superior for branded products. It includes no new calculations, brand dataset or error thresholds. The papers describe research approaches; the proposed acceptance matrix requires adaptation to the actual brief.

Approve perceptual quality and product accuracy separately, then combine those decisions. Provide the project brief with an approved source and a list of features that must remain unchanged. Any metrics used can then serve a clear supporting purpose, while acceptance answers the commercial question: can this particular product be shown in this particular way?