AI Glossary / Computer vision and multimodal AI / Advanced

Covariant derivative (healing brush)

Definition

In the healing-brush context, a covariant derivative is a modified derivative that measures image variation relative to a chosen connection field. Georgiev's formulation uses Dᵢf=∂ᵢf−Aᵢf, with Aᵢ=∂ᵢg/g for a positive reference image g. It provides a mathematical way to compare structure while accounting for a multiplicative reference variation.

What it means

Copying a textured patch directly can leave an obvious brightness mismatch. The theoretical idea is to describe how image values vary relative to a reference, rather than demanding identical absolute values everywhere. For positive g, ∂ᵢg/g equals ∂ᵢlog(g), which explains the connection with ratios and log-domain thinking. That relationship is useful intuition, but it is not a complete specification of a commercial healing tool. The supplied term refers to a particular mathematical interpretation within image editing, not the whole general theory of covariant differentiation.

How it works

Choose a reference field and construct a connection that makes the reference covariantly constant: Dᵢg=0. Use the modified derivative in an image-reconstruction objective together with boundary conditions. Georgiev's historical paper distinguishes this theoretical derivation from a simplified implementation based on reconstructing a difference image with harmonic and biharmonic methods. Consequently, describing the entire historical implementation as merely “take logs and paste gradients” would be too strong. A practical reconstruction also needs discretization, numerical solution, masks and handling of unsuitable reference values.

Three practical examples

• A computational-photography student compares ordinary and reference-adjusted derivatives on a simple positive signal to understand multiplicative variation.
• A research team studies why seamless-cloning objectives can preserve local texture while changing broader brightness relationships, distinguishing the objective from its numerical implementation.
• An editor evaluates a repaired image boundary visually and retains the original, recognizing that a smooth blend is an editing result rather than recovered capture data.

Think of it this way

Imagine comparing prices in two places after accounting for a local conversion factor. Raw price differences mix the item difference with the conversion effect; an adjusted comparison attempts to separate them. The analogy explains why a reference rule changes what counts as “no change.” In image editing the reference is spatial, and the mathematics involves derivatives and boundary conditions. The analogy is not a claim that illumination can always be represented by one perfectly known conversion factor.

Worked example

In an original one-dimensional example, choose positive reference g(x)=e^(2x). Then g′(x)/g(x)=2, so Df=f′−2f. Let f(x)=3e^(2x). Its ordinary derivative is 6e^(2x), but its adjusted derivative is 6e^(2x)−2×3e^(2x)=0. The function is not spatially constant in ordinary coordinates; it is constant relative to this connection because f/g=3 everywhere. This calculation illustrates the operator's meaning. It does not reconstruct an image, select a suitable source patch or establish how a current editing application operates. If the reference were zero, the stated division would require a different treatment.

Illustrative case study

A fictional image-research workshop compares two descriptions of a healing operation. One says it copies pixels; another says it simply performs log-domain cloning. Participants read the historical formulation and build a small mathematical example to see why both descriptions miss distinctions. They separate source texture, reference variation, reconstruction objective and boundary behavior. When assessing an actual edited photograph, they inspect the result without claiming that its appearance proves which internal algorithm was used. The scenario illustrates careful technical interpretation, not a benchmark or a verified account of present-day proprietary software.

Common misconceptions and mistakes

A covariant derivative is not statistical covariance. The chosen connection is part of the model, so zero adjusted derivative does not mean zero ordinary derivative. Another mistake is treating a log-domain identity as a full implementation recipe. Positive-value assumptions matter, and clipping or unsuitable encoding can complicate a direct numerical interpretation. Historical research should also not be presented as complete documentation of every current tool bearing a similar name. Finally, a seamless-looking repair does not establish that the replacement content was present in the original scene.

Business use

This is primarily a technical background term for image-editing research and specialist product discussions. Business users can apply its lesson by separating an editing objective from claims about a tool's internal mathematics. Evaluate the resulting image for the intended communication purpose and keep original material when traceability matters. If a supplier describes a method using this terminology, ask for the actual assumptions and documented implementation before comparing it numerically with another system. A sophisticated mathematical explanation does not by itself establish visual quality or faithful restoration of missing information.

People also ask

Is this the same as covariance between variables?

No. It is a differential operator adjusted by a connection field.

Why does a logarithm appear?

For positive g, the ratio of its derivative to itself is the derivative of log(g).

Does the historical theory fully specify a modern healing brush?

No. The cited paper distinguishes theory and implementation, and later products require their own documentation.

Check your understanding

Why was Df zero in the example?

Answer: The reference adjustment exactly removed the variation shared with g.

Was f itself constant?

Answer: No; only its ratio to g was constant.

What must be supplied to turn the concept into an image-reconstruction procedure?

Answer: An objective, boundary conditions, reference data and a numerical method.

Related terms

• Gradient-domain editing
• connection
• Poisson image editing
• biharmonic reconstruction
• illumination
• seamless cloning

Learn next

Verify Dg=0 algebraically for a positive differentiable reference g. Compare ordinary derivatives with derivatives of log values. Then read the historical paper's theoretical and implementation sections separately, noting where approximations replace the formal derivation.

More in Computer vision and multimodal AI

• Corner detection - Corner detection identifies image locations whose local appearance changes substantially when a small neighborhood is shifted in different directions.
• Correspondence problem - The correspondence problem is the challenge of determining which locations or features in different images refer to the same scene content.
• Coupled hidden Markov model - A coupled hidden Markov model represents several interacting sequences of hidden states, with probabilistic dependencies connecting their evolution over time.
• Cross-dissolve - A cross-dissolve is a transition that gradually reduces one image's contribution while increasing another's.
• Cross / joint bilateral filter - A cross or joint bilateral filter averages values from a target image using weights determined by spatial distance and similarity in a separate guide image.
• Cubic B-spline - A cubic B-spline uses piecewise polynomials of degree three as locally supported basis functions.

Work in finance? Get the free CFO AI Prompt Pack: 7 prompts for finance leaders.

Sources

Georgiev, Photoshop Healing Brush: a Tool for Seamless Cloning: https://ftp.funet.fi/pub/mirrors/ftp.gimp.org/pub/references/Photoshop_Healing_Brush_a_Tool_for_Seamless_Clonin.pdf The exponential calculation, price-conversion analogy and workshop are original educational illustrations. The supplied glossary's ratio-based shorthand needs this theory-versus-implementation qualification.

Previous: Covariance matrix adaptation evolution strategy Next: Cover's theorem