What does the four-part analogy mean?

The transformation from A to A′ is used as guidance for constructing B′ from B.

Does the original framework require a neural network?

No. Its classical approach is based on multiscale neighborhood matching and synthesis.

Can one example handle any new image?

Not reliably. New structures outside the example's coverage can lead to ambiguous or unsuitable synthesis.

AI Glossary / Computer vision and multimodal AI / Advanced

Image Analogies

Definition

Image Analogies is an example-based image-processing framework that uses a pair of images showing an input and its transformed appearance to guide a similar transformation of a new image. The relationship is often written A:A′::B:B′: given A, its processed version A′ and a new input B, synthesize an analogous output B′.

What it means

Instead of specifying every filter mathematically, a user can demonstrate the desired relationship through an example pair. A photograph and a corresponding painted version might illustrate a rendering style, while a label image and textured result can illustrate a mapping from regions to appearance. The named framework introduced by Hertzmann and colleagues uses multiscale neighborhood matching and synthesis. It is historically distinct from modern neural style-transfer methods, although both explore example-guided transformation. The analogy is visual and statistical; the method does not infer a universally correct abstract rule from one pair.

How it works

The method builds multiresolution representations of the example pair and new input. It compares neighborhoods in the new image with neighborhoods in the example, considering both input appearance and already synthesized output context. Corresponding information from A′ guides the output B′. Processing from coarse to fine helps establish broad structure before adding detail.
The example pair must provide a meaningful relationship, usually with appropriate correspondence between A and A′. Nearest-neighbor matches and spatial coherence influence synthesis. If the new image contains structures unlike the example, the inferred transformation may fail or repeat unsuitable patterns. The result depends strongly on the examples and representation, not only on the broad task label.

Three practical examples

A graphics researcher demonstrates an artistic rendering effect through a source image and its stylized counterpart, then applies the relationship to another source.
A designer uses a texture-by-numbers setup in which labeled regions guide the placement of example textures.
A computer-graphics course compares example-based filtering with an explicitly programmed filter to explore what an example pair does and does not specify.

Think of it this way

Think of showing a craftsperson an unfinished sample and the same sample after treatment, then asking for a similar treatment on a new object. The example communicates a relationship but may leave important choices ambiguous. The analogy is limited because the algorithm matches image neighborhoods; it does not possess the craftsperson's general understanding of materials, purpose or artistic intent.

Worked example

Suppose A is a grayscale drawing of a simple house and A′ is the same drawing rendered with brick-like walls and a textured roof. B is a new house drawing with a different arrangement. The framework seeks neighborhoods in A resembling parts of B and uses their corresponding appearance in A′ to guide B′. A wall-like region can therefore receive wall-like texture while respecting available local context.
If B contains a glass dome absent from A, the example offers no clear analogous treatment. The algorithm may borrow an unsuitable nearby pattern. This is a worked reasoning example, not a reproduced experiment or a claim that the system recognizes houses semantically. It illustrates dependence on the example's coverage.

Illustrative case study

Fictional example: Juniper Illustration tests an example pair to turn clean line drawings into a textured poster style. Results on similar drawings are useful, but a new image with dense lettering receives texture that reduces readability. The team separates the text layer and revises the example or workflow. It keeps the original inputs and generated output distinct. The case illustrates that a successful visual analogy on one class of content does not establish a universal transformation rule.

Common misconceptions and mistakes

Image Analogies is not simply copying the whole example onto the new image. It is not the same as a text-to-image generator or a modern neural style-transfer model. An example pair does not uniquely define the intended transformation in every possible situation. Fine local matches can still produce globally unsuitable results. More texture is not automatically a better analogy, especially for text or precise diagrams. Finally, synthesized appearance should not be treated as a recovered physical property of the original scene.

Business use

Businesses exploring example-guided design should select examples that cover the intended content and review where they do not. Protect text, logos and product details that must remain exact. Compare output consistency across a representative set rather than approving one attractive demonstration. Preserve source pairs and settings for reproducibility, and use example material with suitable rights for the intended output. The technique is useful for controlled creative exploration when its dependence on examples and potential failure modes are understood.

Check your understanding

An example pair contains only smooth landscapes, but the new image contains small technical text. Is readable text output guaranteed?

Answer: No. The example may not specify a suitable transformation for that structure, so the text needs separate protection or another workflow.

Related terms

Texture synthesis
Image quilting
Nearest neighbor search
Non-photorealistic rendering (NPR)
example-based rendering
style transfer
image pyramid
non-photorealistic rendering

Learn next

Learn local image neighborhoods and multiscale representations, then examine how an aligned before-and-after pair constrains a transformation. Compare supported and unsupported input structures before studying the original synthesis procedure.

Sources

Hertzmann and colleagues, Image Analogies, authors' publication page: https://mrl.cs.nyu.edu/publications/image-analogies/ Primary paper: https://www.cs.jhu.edu/~misha/ReadingSeminar/Papers/Hertzmann01.pdf Primary publication page and paper excerpts consulted; full implementation and editorial review remain pending. The house and poster scenarios are original illustrations.

Previous: Image Next: Image-based lighting