Full Photometric Model Overview

• Full Photometric Model is an end-to-end framework that converts scene radiance into measured brightness by modeling internal camera effects and external contaminants.
• It integrates sensor nonuniformity and vignetting with factors like smudges and defocus, enabling comprehensive reconstruction of photometric measurements.
• Applications span optical surveys, projector compensation, and satellite brightness prediction, yielding measurable improvements in image quality metrics.

A full photometric model is an end-to-end description of how an underlying radiometric, scene, or surface quantity is transformed into a recorded photometric measurement. In the 3D scene-camera formulation, it explicitly factors internal camera response, external contamination, and defocus into a single image-formation model and jointly optimizes those terms with the scene representation (Dai et al., 26 Jun 2025). In adjacent literatures, the same idea appears as a complete system-sensitivity model for survey photometry, a coupled photometric-geometric forward model for projector compensation, and a BRDF-based brightness model for satellites (Lorenzo-Gutiérrez et al., 2019, Li et al., 2 Apr 2026, Lu, 2024). The unifying feature is not a single canonical equation, but the requirement that all dominant transformations between source and measurement be represented within one coherent framework.

1. End-to-end scope and formal structure

In optical survey photometry, a full model is often anchored in the total system sensitivity curve. For GALANTE, the fraction of stellar flux recorded on the detector at wavelength $\lambda$ is written as

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$

where the factors are the filter transmission, CCD quantum efficiency, telescope-mirror reflectivity, and atmospheric transmission, with $A(\lambda)$ computed at an airmass of $1.3$ (Lorenzo-Gutiérrez et al., 2019). Synthetic AB magnitudes are then obtained from the filtered SED through

$$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$

ALHAMBRA adopts the same structural logic, expressing the total response as

$$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$

with atmospheric transmission evaluated at $1.2$ airmasses (Villegas et al., 2010). In both cases, first- and second-order moments of the response curve are used to characterize the effective spectral location and spread of each band.

A complementary standardization of full photometric description is provided by the IVOA Photometry Data Model. There, a usable photometric calibration is not just a filter curve, but the bound combination of a PhotometryFilter, a MagnitudeSystem, and a ZeroPoint through the PhotCal entity (Salgado et al., 2014). This formalization is designed to permit conversion from magnitudes to physical flux densities and to expose filter profiles, zero points, and calibration metadata to clients and archives. In that sense, a full photometric model also includes the metadata and reference-spectrum conventions needed to interpret a measurement physically.

At the detector level, the most expansive formulation in the provided material is the Dark Energy Camera model, which maps sky brightness to raw ADU through collected electrons, brighter-fatter charge redistribution, amplifier nonlinearity, gain, crosstalk, and residual additive backgrounds (Bernstein et al., 2017). This makes explicit that a full photometric model may extend beyond bandpass definition into the full instrument-response chain.

2. Internal and external factors in the scene-camera model

In the scene-camera representation with joint camera photometric optimization, the full photometric model is divided into internal and external components (Dai et al., 26 Jun 2025). The internal photometric model captures two spatially varying multiplicative effects on the incoming radiance $\mathbf B_i(\mathbf x)$ at pixel $\mathbf x$ in image $i$: vignetting $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$0 and sensor response or non-uniformity $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$1. The recorded image is therefore

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$2

and the product

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$3

is used as the internal-photometric term.

The external photometric model is introduced to represent surface contamination such as smudges, dirt, and water droplets. In that formulation, contamination both attenuates scene radiance and adds an emissive term:

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$4

where $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$5 is a local attenuation factor, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$6 is an additive emissive term, and $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$7 is the ideal scene radiance rendered from the 3D model (Dai et al., 26 Jun 2025). The decomposition is intentionally asymmetric: internal effects are multiplicative and camera-centric, whereas external effects are modeled as attenuation plus additive glow located in image space.

Defocus is then incorporated through a finite-depth-of-field model. A point at object distance $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$8 produces a circle of confusion with radius

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$9

where $A(\lambda)$0 is focal length, $A(\lambda)$1 is aperture diameter, and $A(\lambda)$2 is focus distance (Dai et al., 26 Jun 2025). The external model is averaged over the CoC, so the full model under defocus becomes a spatial average of $A(\lambda)$3 multiplied by $A(\lambda)$4. When one assumes that scene points lie within the depth of field so that the CoC collapses to $A(\lambda)$5, the simplified 2D-2D expression is

$A(\lambda)$6

This construction makes the phrase “full photometric model” more specific than a generic exposure or gain correction. It denotes a factorization of image formation into scene-related radiance and scene-unrelated camera effects, with explicit accommodation for blur and contamination rather than absorbing them implicitly into the radiance field (Dai et al., 26 Jun 2025).

3. Parameterization, joint optimization, and depth regularization

The practical parameterization in the scene-camera method uses two shallow MLPs to represent the three photometric maps $A(\lambda)$7, $A(\lambda)$8, and $A(\lambda)$9 (Dai et al., 26 Jun 2025). The paper denotes $1.3$0 as learning the internal-photometric term $1.3$1, while $1.3$2 and $1.3$3 model the external terms $1.3$4 and $1.3$5. Scene parameters $1.3$6 are the 3D Gaussians $1.3$7 together with opacities $1.3$8 and color coefficients $1.3$9, whereas camera photometric parameters $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$0 are the MLP weights.

Optimization is driven by a photometric reconstruction loss,

$$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$1

with the total loss taken as $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$2 plus any small regularizers if used (Dai et al., 26 Jun 2025). The learning schedule alternates between camera-only and scene-only updates: first fixing the scene and optimizing camera MLPs over a batch of views, then fixing the camera and optimizing the scene Gaussians by splatting-based gradients. The paper states that this schedule ensures that the scene model does not “explain away” photometric effects and vice versa.

A specific risk in such factorized models is that the photometric branch absorbs geometry, producing floating “ghost” Gaussians or incorrect depth assignments. The method addresses this with a depth regularization applied during camera-MLP updates. Gaussians whose inverse depth is more than $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$3 away from the mean along the ray are downweighted through a soft Gaussian gate

$$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$4

with the rendered radiance modified accordingly (Dai et al., 26 Jun 2025). This is a regularization against photometric overfitting, not a replacement for geometric supervision.

The reported results indicate that the approach improves robustness under image degradation. On mixed real-world distortions, the averages reported against 3DGS-MCMC are: fingerprint, $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$5 versus $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$6 in PSNR/SSIM; dirt, $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$7 versus $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$8; water droplet, $$m_{\rm AB}=-2.5\log_{10}\!\left[\frac{\int f_\lambda S(\lambda)\lambda\,d\lambda}{\int S(\lambda)\lambda\,d\lambda}\right]-48.60.$$9 versus $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$0; and vignetting, $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$1 versus $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$2 (Dai et al., 26 Jun 2025). On Sony A7R3, Logitech webcam, and OPPO phone data, improvements of $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$3–$$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$4 in PSNR and $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$5–$$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$6 in SSIM over the 3DGS baseline are reported. The ablation study states that adding the camera photometric representation gives $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$7 dB, while adding depth regularization or the defocus model adds approximately $$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$8–$$R(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times T_{\rm atm}(\lambda),$$9 dB each, and all together yield $1.2$0 dB over baseline (Dai et al., 26 Jun 2025).

4. Calibration-oriented full photometric models in survey systems

In survey photometry, a full photometric model typically combines throughput definition, characteristic band descriptors, transformations to reference systems, and a zero-point strategy. GALANTE exemplifies this integrated pattern (Lorenzo-Gutiérrez et al., 2019). Its seven intermediate- and narrow-band filters cover $1.2$1 to $1.2$2 and are characterized both by the total sensitivity curve and by moments such as

$1.2$3

The filter set consists of F348M, F420N, F450N, F515N, F660N, F665N, and F861M, with quoted isophotal wavelengths and effective widths. Transformations from SDSS $1.2$4 to GALANTE magnitudes are given as one-colour linear relations of the form

$1.2$5

with specific relations and rms residuals reported for all seven filters. For example, $1.2$6 with $1.2$7, while $1.2$8 with $1.2$9 (Lorenzo-Gutiérrez et al., 2019).

GALANTE zero-point calibration was first derived using SDSS DR12 and then repeated with RefCat2 for griz, retaining SDSS DR12 $\mathbf B_i(\mathbf x)$0 for the F348M transformation because RefCat2 lacks a $\mathbf B_i(\mathbf x)$1 band (Lorenzo-Gutiérrez et al., 2019). The zero point in each band was taken as the median of $\mathbf B_i(\mathbf x)$2 over approximately $\mathbf B_i(\mathbf x)$3 stars, and the resulting RefCat2-based zero-point distributions were described as significantly narrower and more symmetric than those based on DR12. The adopted calibration is

$\mathbf B_i(\mathbf x)$4

The paper also emphasizes caveats: atmospheric extinction is folded into $\mathbf B_i(\mathbf x)$5 at $\mathbf B_i(\mathbf x)$6 airmasses; F348M is especially sensitive to the Balmer-jump area and has somewhat larger residuals; DR8 suffered from an “uber-cal” magnitude equation in $\mathbf B_i(\mathbf x)$7 and $\mathbf B_i(\mathbf x)$8 that was corrected in DR12; and future work will refine the preliminary zero-point calibration with an internal network of spectrophotometric standards observed directly with GALANTE (Lorenzo-Gutiérrez et al., 2019).

ALHAMBRA provides a closely related formulation at larger filter multiplicity. It uses a 20 contiguous, equal-width, medium-band CCD system from $\mathbf B_i(\mathbf x)$9 to $\mathbf x$0, characterizes the optical response by first and second moments, and defines its AB scale using 31 classic spectrophotometric standards together with 288 stars from the Next Generation Spectral Library (Villegas et al., 2010). The transformation between SDSS and ALHAMBRA is expressed as a linear combination of nearby SDSS colours, and zero points are obtained by matching field stars to best-fitting NGSL standards, assigning synthetic ALHAMBRA AB magnitudes, rejecting outliers with a Chebyshev filter, and adopting the median residual as $\mathbf x$1 (Villegas et al., 2010).

These systems show that, in observational photometry, a “full” model often means more than throughput. It includes the response curve, representative wavelengths, synthetic-photometry formalism, cross-system transformations, and a reproducible zero-point procedure.

5. Coupled photometric-geometric and BRDF-based formulations

A different use of the term arises in projector compensation, where the photometric model is explicitly coupled to geometry. In SIComp, the complete forward process is

$\mathbf x$2

where $\mathbf x$3 is the RGB value sent to the projector, $\mathbf x$4 is the per-pixel surface reflectance, $\mathbf x$5 is ambient illumination, $\mathbf x$6 is the unknown nonlinear photometric transfer, $\mathbf x$7 is the geometric warp, and $\mathbf x$8 is the RGB value captured by the camera (Li et al., 2 Apr 2026). The ideal inverse seeks a compensation image $\mathbf x$9 such that

$i$0

but because neither $i$1 nor $i$2 is analytically invertible in closed form, SIComp approximates the inverse with an optical-flow-based geometric inverse and a CNN-based photometric inverse. The photometric network is further conditioned on intensity-varying surface priors produced by projecting gray levels $i$3 and capturing the corresponding warped priors. The assumptions are explicit: surfaces are treated as nearly Lambertian with spatially varying reflectance; ambient light is additive; projector-camera color mapping is deterministic but unknown; and no specular or interreflection terms are explicitly modeled (Li et al., 2 Apr 2026).

The SIComp experiments distinguish surrogate digital testing from real projection, and report PSNR, RMSE, SSIM, $i$4, LPIPS, and FID. On Set A, the best SIComp variant with $i$5 reports $i$6 PSNR, $i$7 RMSE, $i$8 SSIM, $i$9 $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$00, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$01 LPIPS, and $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$02 FID; on Set B, SIComp$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$03 reports $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$04 PSNR, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$05 SSIM, and $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$06 $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$07 (Li et al., 2 Apr 2026). The paper states that going from $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$08 to $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$09 intensity priors yields the largest gain and that $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$10 gives a further modest improvement.

In BRDF-based satellite photometry, a full model is again end-to-end, but the controlling variables are attitude, geometry, and facet reflectance rather than camera contamination. The Starlink V1.5 model treats the satellite as two planar facets, a solar array and a chassis, each with diffuse and glossy/specular terms, plus a sail back-scattering term, chassis blocking factors, and an earthshine component (Lu, 2024). The total reflected power into the observer’s line of sight is written as the sum of sail specular, sail diffuse, sail back-scatter, chassis specular, chassis diffuse, and earthshine contributions. Apparent Johnson-$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$11 magnitude at normalized range $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$12 km is then

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$13

The fitted parameters from MCMC are $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$14, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$15, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$16, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$17, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$18, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$19, and $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$20, with predictions reported to within approximately $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$21 mag rms (Lu, 2024).

Taken together, these formulations show that a full photometric model is domain-dependent in its parameterization but consistent in ambition: it attempts to represent the entire forward path from source or desired input to measured brightness.

6. Probabilistic interpretation, standardization, and recurring limitations

Full photometric modeling also appears in probabilistic inference, where the forward photometric relation becomes the basis for posterior estimation rather than direct correction. XDQSOz models the joint density of quasar relative fluxes and redshift as a Gaussian mixture in $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$22 space, learned by extreme deconvolution under heteroscedastic uncertainties and missing data (Bovy et al., 2011). Classification probabilities, redshift-slice probabilities, and the full redshift posterior follow analytically from the conditioned mixture. The reported impact of added UV and NIR photometry is strong: for spectroscopically confirmed SDSS quasars with GALEX UV and UKIDSS NIR data, $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$23 percent have photometric redshifts within $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$24 and $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$25 percent within $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$26 of the spectroscopic redshift, described as about a factor of three improvement over ugriz-only photometric redshifts (Bovy et al., 2011).

A newer generative formulation infers both redshift and rest-frame spectra from broadband fluxes using a latent-diffusion model conditioned on photometric fluxes $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$27 (Guo et al., 11 May 2026). Its forward likelihood is

$$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$28

with Gaussian photometric noise in each band. Sampling from the conditional diffusion model yields a photometric-redshift PDF as well as latent spectral codes that decode to rest-frame spectra. For galaxies with sufficiently high signal-to-noise ratios in the observed spectra, the reconstructed and observed Dn4000 indices show Pearson $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$29 (Guo et al., 11 May 2026). This suggests that, in probabilistic settings, a full photometric model may be defined less by explicit detector physics than by a complete generative specification linking latent spectra, redshift, filter throughput, and observed fluxes.

Several recurring limitations cut across these literatures. In the scene-camera model, the defocus model is described as crude, with a single uniform CoC and no depth dependence for scene points; over-parameterized MLPs can still absorb some geometry; the method does not model per-channel color response curves or nonlinear gamma; and it assumes static scene and static camera parameters per dataset (Dai et al., 26 Jun 2025). In SIComp, no specular or interreflection terms are explicitly modeled, and failures under strong specularities or occlusion are discussed as limitations (Li et al., 2 Apr 2026). In GALANTE, atmospheric extinction and catalogue-based systematics remain explicit caveats, and the F348M band is unusually sensitive because RefCat2 lacks a $$S(\lambda)=T_{\rm filter}(\lambda)\times QE(\lambda)\times R(\lambda)\times A(\lambda),$$30 band (Lorenzo-Gutiérrez et al., 2019). In DECam, achieving final uniformity requires not only flat-fielding but also sky and fringe subtraction, brighter-fatter correction, separation of pixel-area and QE effects, and pixel-by-pixel spectral-response characterization (Bernstein et al., 2017).

A common misconception is that a full photometric model is synonymous with a filter transmission curve or a zero point. The literature summarized here does not support that reduction. In every domain represented, “full” denotes a model that spans the complete forward measurement chain relevant to the target task: throughput and calibration in survey photometry, contamination and blur in scene-camera reconstruction, geometry-coupled inversion in projector compensation, BRDF and attitude in satellite brightness prediction, or latent-spectrum and redshift inference in probabilistic photometric analysis (Salgado et al., 2014, Dai et al., 26 Jun 2025).