Skip to content

Latest commit

 

History

History
41 lines (25 loc) · 2.53 KB

tips.md

File metadata and controls

41 lines (25 loc) · 2.53 KB
Raw

TLDR

  1. Manually reduce the input image size and upscale twice to get better results
  2. Don't be afraid to add noise or blur the input image to potentially improve results
  3. For NEXT, set the strength of the conditioning to 1, detail to 1, stars to 0.

Conditioning

Experiments with NEXT and conditioning are ongoing. My current advice is to set the strength of the conditioning to 1, details to 1 and stars to 0, however, I typically see that trying to remove stars also removes some detail. I also am not sure if I will keep/update these embeddings for better alternatives, or (more likely) add more options. For now though, this is mostly unexplored for you to experiment on a per-image basis.

Crushing

The model that tries to improve the image is trained on really noisy and blurry inputs. If the input image is too clean, you may get artifacts and sub-par performance. That is why I recommended making the image worse by manually adding blur/noise/downscaling with an image processing tool such as gimp or photoshop.

The most significant factor that determines the quality of the output, is the input size. After reducing image size such that the stars are at no more than 6pixels wide, performance can drastically improve.

Another factor is the presence of diffraction spikes, as these are rarely present in the training data and as such, are more "troublesome" and can get a little wavy or split up into smaller stars. I unfortunately have to recommend not having diffraction spikes in the input image to begin with.

Workflow for difficult cases

  1. Original -> +downsample 4x -> +noise -> +blur -> "Crushed" image
  2. "Crushed" image -> +upscale -> +upscale -> Output

Workflow result on original/sample2.png

Raw AstroSleuthV2 Workflow with V2 Workflow with V1

As you can see, this workflow works much better with AstroSleuthV1 for this particular image. However I find results can drastically differ based on the amount of noise/blur applied to the crushed image, but were kept the same for the sake of comparison.

Image credits / sources

https://www.reddit.com/r/astrophotography/comments/12fnhb3/m35_and_ngc_2158/ https://www.reddit.com/r/astrophotography/comments/12flzf1/messier_106_in_lrgb/ https://www.reddit.com/r/astrophotography/comments/12febjp/orion_nebula/ https://www.reddit.com/r/astrophotography/comments/12fewff/northwestern_cygnus_nebulae/