DeepPRIME grid artifacts at insane settings

An illustration of some DeepPRIME (DP) internals.
Probably unrelated to real life, just for fun.
Maybe related to DeepPrime XD weird artifacts .

The following two pictures were generated from the same RAW file using DP and DPXD respectively:
Nikon D780, f/16 @105mm, 1/50s, ISO 204k (Hi2.0, two stops above the nominal max. ISO).
Photo of a grey card lit by 2700K LED bulb.

  • Rendering: not active (my default is ‘No corrections’)
  • Microcontrast = +100
  • Tone Curve: bottom sliders set at 105 and 115, R and B channels turned off.
  • Denoising: Luminance = +100, Noise model = -100 for both DP and DPXD versions.
  • Export: quality=30%, size=2Mpx (to make output smaller, no big difference at q=100%)
  • DPL7.3 (+ FP7.3 + VP4.13)

One could also add R and B channels and set ClearView=100, but I wanted to keep it simple. The unprocessed photo was full of noise, 99% photon, very little noise coming from the box.

Same results were obtained on two computers:

  • i7-14700KF (12 E-cores, 8 P-cores), RTX 4070 (56 SMs), Win11
  • i3-10110U (2 cores), no dedicated graphic card, Win10.

Hence it’s not related to particular CPU or GPU.

Sensor is 6048x4024 pixels.
The grid is formed by 1024x1024 pixels for pure DP, and 416x416 pixels for DP XD.
It didn’t show when using HQ or PRIME.
I was diverted to this example while figuring out the noise generated by the box itself.

nice catch …

here is raw → DxO PL7.3 w/ no corrections preset + then DPXD only enabled → linear DNG (only NR & optics ) → ACR and adjusted in ACR to illustrate the blocks

1 Like

the same also with more clearly illustrated DxO maze demosaicking artefacts

1 Like

I’ve reproduced this “feature” for Canon R5 (8192x5464) and Nikon Z8 (8256x5504).
The results are the same as for 24MPx D780: DPXD “blocks” are 416x416,
while DP blocks are 1024x1024. So its probably DxO’s compromise between quality
and performance. The feature is seen only at high DeepPRIME.Luminance settings which
nobody sane would use. Hence, to simplify things and maybe for performance reasons (?), DxO didn’t use overlapping regions.

For example, take
Canon EOS R5 sample gallery: Digital Photography Review
and set

  • DeepPRIME.Luminance = 100
  • DeepPRIME.Noise model = -100
  • Exposure = +1.0
  • Microcontrast = +100

This photo has unexpectedly strong electronic artifacts for an ISO 50k shot,
but nevertheless the blocks are clearly visible.

As for the maze, it can be amplified by ‘Unsharp mask’.
Using also other sliders for the “noise photo”, some interesting textures
can be produced, for which I have yet to find application :wink:
I was trying to get some non-typical random grain patterns.
The grain provided by DxO probably comes from replicating the scan of 7x7mm
film area and the repeating pattern can be quite visible for some films.
Perhaps ‘Agfa APX 100’ and ‘Kodak Tri-X 400’ are uniform enough to use safely.

that was discussed before - DxO just repeats the same small patch in a regular checkerboard order for grain emulation w/o any slightest attempt to randomize application …