The Ansel-nn-parity binary checks that the neural network behind the AI raw denoise module computes the same thing on your CPU and on your graphics card as it did on the machine that trained it.

You should not normally need it. It exists for one specific situation: the denoised image looks wrong — a grid, a mesh, a colour cast — and you suspect the graphics card. This tool answers that question with a number instead of an opinion, and tells you which of the three implementations disagrees with the other two.

Running it 

1Ansel-nn-parity <model.anselnn> <fixture-directory> [size] [--core <Ansel options>]

The model is one of the .anselnn files shipped with Ansel (typically in /usr/share/ansel/ or share/ansel/ inside your installation). The fixture directory holds a reference input and the output the training framework produced from it; fixtures live in the ansel-denoise  repository under fixtures/models/, one per model, and must match the model you point at — the tool refuses to run otherwise, because comparing one model against another model’s reference produces a large error that looks exactly like a real defect.

Anything after --core is passed to Ansel itself, which is how you select a device:

1Ansel-nn-parity share/ansel/denoise-large-single-v1.anselnn fixtures/models/large-single 96 \
2    --core --conf opencl_devid_export="+1"

Reading the output 

1model denoise-large-single-v1.anselnn: in=5 out=1, fixture 96x96
2  torch vs CPU     : max abs err 1.79e-07 (at 440: 0.535022 vs 0.535022)
3  using OpenCL device 0
4  torch vs OpenCL  : max abs err 1.79e-07 (at 440: 0.535022 vs 0.535022)
5  CPU   vs OpenCL  : max abs err 1.19e-07
6  PASS (tolerance 2e-04)

Healthy numbers sit around 1e-07, the precision limit of single-precision arithmetic. The tolerance is 2e-04, so a passing result has roughly a thousandfold margin.

What a failure tells you depends on which line fails:

  • torch vs OpenCL fails while torch vs CPU passes — your graphics driver is the problem, not Ansel and not the model. The usual cause is an inaccurate math library in the driver; see possible problems & solutions.
  • Both fail by the same amount — the fixture does not belong to that model. Check that you paired them correctly.
  • torch vs CPU fails alone — that would be a genuine bug in Ansel. Please report it.

If no OpenCL device is available the GPU comparison is skipped and only the CPU is checked.

What it does not cover

The reference input is synthetic, so the tool validates the network and the OpenCL kernels — not the whole module. A correct result here means the arithmetic is sound; it does not prove that every rendering path around it is. It is a hardware and driver check, not a substitute for looking at your pictures.