Seiza’s first published noise model, noise-v1-d33, works on linear FITS and XISF stacks. It estimates what to remove from the image you give it, so you can use a stack made in another program. Two controls let you choose how much fine grain and wider mottle to remove. Seiza 0.52 release notes

Make two stacks from separate exposures

We train on pairs of images made from the same target. Seiza stacks a small group of exposures for the input and a deeper, separate group for the target. Both stacks use the same reference grid. We interleave the groups through the session to keep seeing, airmass and gradients comparable.

The deeper stack still contains noise, but the two stacks share no exposures. The stars and nebula appear in both, while random noise varies between them. That gives the model a way to learn from real data, including the effects of resampling and rejection. Shared calibration patterns can appear in both stacks, so the assumption has limits.

The two stacks also need to agree on the light they recorded. A change in transparency can shift the flux scale, and haze can spread extra light into the wings of stars. We match the pairs’ flux and reject pairs whose star cores or wings still disagree. We group overlapping fields before splitting the data for training and validation so crops from the same sky don’t appear on both sides. Training design

Check drizzled stacks and calibration defects

Drizzle spreads some of a pixel’s noise into nearby pixels. A noise estimate based on local pixel differences can then read low because neighbors move together. Seiza measures that correlation and raises the noise estimate when it exceeds the level the model trained on. Seiza 0.51 notes

Deep dark defects caused a different failure. A dark frame can subtract too much from a hot pixel and leave a hole below the sky level. Early runs filled those holes and sometimes overshot into bright spikes. The runtime now detects unusually deep holes and temporarily fills them with nearby values before the model runs. Afterward it restores the affected input values where the output differs enough to trigger the guard. On one drizzled Hα stack of the Elephant’s Trunk, the largest change inside a guarded hole fell from 6,858 to 15 times the image’s estimated noise level. That test measures how much the model changed the defects. The original defect still belongs in calibration and rejection. Training record

Adjust fine grain and wider mottle

Seiza splits the model’s proposed correction into fine and broad bands. Set how much of each correction to apply. A lower Grain setting keeps more of the fine texture in the original image. A higher Mottle setting removes more of the broader, patchy noise. Compare the result at the same stretch and check faint stars and nebula edges as well as empty sky.

In PSF Guard 0.14, install the model under Settings › Processing › Stacking › Processing tools. Choose Seiza noise model under Noise reduction when processing a stack. In the Seiza CLI, use seiza model denoise with --high-amount for grain and --low-amount for mottle. PSF Guard release notes · Seiza CLI guide