Scaling contract (MED-28): the photon-gain term scales as DPRScale²
while the readout and FPN sigmas scale linearly with DPRScale.
This asymmetry is intentional and reflects the underlying physics:
photon-gain models shot noise, whose variance is proportional to
the number of samples integrated per output pixel (≈ DPRScale²).
The gain term enters the shader as a variance, so it scales by
DPRScale².
readout and FPN sigmas are per-pixel noise terms independent of
integration time / sample count, so they scale linearly with
DPRScale to preserve the visual standard deviation as the
effective pixel size changes.
Do NOT "fix" this for visual symmetry — squaring or linearising
both terms breaks the noise model. When DPR < 1 the effective
pixel is larger, so all three terms shrink (gain quadratically,
sigmas linearly). When DPR > 1 they grow in the same proportions.
Re-apply base sigmas through the DPR scaling.
Scaling contract (MED-28): the photon-gain term scales as
DPRScale²while the readout and FPN sigmas scale linearly withDPRScale. This asymmetry is intentional and reflects the underlying physics:DPRScaleto preserve the visual standard deviation as the effective pixel size changes.Do NOT "fix" this for visual symmetry — squaring or linearising both terms breaks the noise model. When
DPR < 1the effective pixel is larger, so all three terms shrink (gain quadratically, sigmas linearly). WhenDPR > 1they grow in the same proportions.