title: Intrinsic Decomposition: Assumption and Review Contract description: "Intrinsic decomposition is an ambiguity-bound estimate of reflectance and illumination, not ground truth; bind the image-formation assumptions, model and version, color domain, residual policy, source evidence, and task-specific review before using its albedo or shading outputs for editing or relighting." category: techniques tags: [intrinsic-decomposition, albedo, illumination, reflectance, shading, relighting, evidence] aliases: ["Albedo and Illumination Separation", "Intrinsic Image Decomposition"]
Intrinsic Decomposition: Assumption and Review Contract¶
Intrinsic decomposition estimates components such as reflectance (often called albedo) and shading from an image. It is not a measurement of the material or lighting in a photographed scene. In a simplified diffuse model, an observed image can be expressed as I ≈ R ⊙ S + E, where R is reflectance, S is illumination-dependent shading, and E absorbs effects the model does not represent. Real images include specularity, transparency, subsurface scattering, sensor processing, clipping, and mixed light. A single RGB image usually does not uniquely determine these components.
Declare the decomposition contract¶
For every output, retain:
- source asset digest, crop/orientation, color encoding, alpha policy, and whether the model sees linear, scene-referred, or display-referred values;
- exact model release, weights, runtime, prompt/reference inputs, and the image-formation assumptions the method documents;
- named outputs, their units/normalization, residual or confidence handling, and which pixels have no reliable estimate;
- intended operation, such as a lighting-only preview or a constrained recolor, plus protected content; and
- comparison source, numerical/reconstruction check where meaningful, and delivery-resolution visual review.
An albedo-like output from one method cannot be substituted for another method's albedo, normal, material, or relighting input without an implementation-specific compatibility test.
Treat estimates as model-scoped¶
The task spans different observations and assumptions. For example, Intrinsic Scene Decomposition From RGB-D Images uses depth and a stated illumination model, while Intrinsic Image Transformation via Scale Space Decomposition frames single-image outputs as learned image-to-image estimates. Those papers do not establish that a named output is physically correct for arbitrary RGB photographs, skin, products, transparent objects, or a later model release.
Select a method only after testing the exact release on representative images and the downstream operation. Record failure cases such as hard shadows, specular highlights, clipped lights, saturated colors, mixed illumination, thin structures, and camera-processing artifacts.
Use a constrained edit workflow¶
- Verify that the source, rights, crop, color domain, and protected regions are within the job scope.
- Produce the stated estimate and keep the source-to-output association.
- Inspect the residual and ambiguity-prone regions before treating a value as reflectance or shading.
- Apply only the approved change to explicit masks. A low-frequency or albedo-like layer is not permission to change every material or face pixel.
- Recompose only with the model's declared routine. If a generated image, clipping, blend, or nonlinear step is involved, label the result as derived rather than exact reconstruction.
- Review source versus result at delivery resolution for color drift, halos, shadow/material swaps, detail removal, geometry changes, and protected content.
Albedo-domain color transfer can be a useful experiment when illumination confounds a color edit. It still needs a source-aware comparison and does not prove that the target color is physically correct. Preserve a residual only when the selected method defines it and the output contract says how it is combined.
Failure and release boundary¶
Do not silently replace an uncertain estimate with a different model, add generated texture, or present a relit result as recovered source data. Return a visible review or failure state when assumptions are unknown, the runtime cannot reproduce the declared output, the residual is unexplained, or review finds a protected-content change.