Documentation
This section collects technical notes and usage guides for StarNet and DeepSNR: command-line references, input behavior, data-quality guidance, and workflow examples.
DeepSNR input guidance
This page describes DeepSNR input compatibility, channel behavior, and input issues that can cause artifacts.
General input data compatibility notes
DeepSNR works best when the image statistics match the training data, which had mostly uncorrelated high-frequency noise. Think about compatibility in two parts: where the data came from and how the channels are represented.
Data origin
DeepSNR is intended for images from monochrome cameras. Those images can remain grayscale or be combined into color images, including narrowband combinations such as SHO.
One-shot color (OSC) images can also work well when they have been drizzle-integrated. That path gives the model cleaner per-channel image structure than a simple debayered stack.
Color and channel support
Model/weights v1 is the older RGB-only path. It is not a good fit for grayscale stacks or color images with identical or highly correlated channels, such as some HOO combinations.
Model/weights v2 is the current preferred path. It supports grayscale and color images, and it handles images with two highly correlated channels better than v1.
Data quality
DeepSNR is intended for high-quality calibrated and integrated data where the remaining noise is mostly random and not strongly patterned.
Noise that is already baked into the image as correlated structure is much harder to remove cleanly. Walking noise is a typical example: because it behaves like image structure rather than random noise, DeepSNR should not be expected to fix it well. Correct acquisition, calibration, dithering, and integration problems before using DeepSNR.