uberSmooth: AI Deep Astrophotography Blind Image Denoising and Sharpening App
Version 0.1 now available:
There are similar quality tools on the commercial market, so I'm asking for a donation of *any* amount to help cover costs during development and model training. (Entering an amount under $1.00 USD will provide gratis download without any payments or accounts required to avoid processing fees). I plan to eventually release the model training source code and encourage community sharing of trained and fine-tuned models for everyone to use. If you would like custom trained models, hands on help processing your data, or access to training code now, please contact me via my email listed below to discuss possible arrangements. This uberSmooth inference/prediction CLI binary is released under the apache-2.0 license. The two model weight files included with the binary are released under the cdla-sharing-1.0 license, and are both fine-tunes of the original SCUNet paper model, upon which the uberSmooth v0.1 architecture is also based. All dependencies and libraries used in this software are subject to their respective licensing agreements and copyright notices as specified by their original authors. Special thanks and acknowledgments to NASA and STScI for providing high resolution image data used in training from their webbtelescope.org and hubblesite.org sites which may be freely used as in the public domain in accordance with NASA's contract. This version of uberSmooth currently works best on fully processed and stretched non-linear images. Simpy use your normal workflow but limit or entirely skip denoise and deconvolution steps. Run uberSmooth as the last step before publishing. It may work okay on other data as well. It is compiled with NVidia cuda support. My old NVidia 1070 GTX takes under ~2 minutes to process a full APS-C 25MP 6k x 4k resolution image. It may take over ~20 minutes to process if only CPU is available. Supported file formats include:
Images, logos, and uberSmooth name Copyright © 2024 John W. Leimgruber III (ubergarm) unless otherwise credited.
- Windows release v0.1 is built and available for download below!
- Linux release v0.1 is built (but not yet available for download)
- Mac release v0.1 coming soon :TM:
usage: uberSmooth [-h] [-V] -i INPUT_FILE -o OUTPUT_FILE -m MODEL_FILE [-d DEVICE] [-b BLEND] [-v]
options:
-h, --help show this help message and exit
-V, --version display version string and exit
-i INPUT_FILE, --input_file INPUT_FILE
Input file path
-o OUTPUT_FILE, --output_file OUTPUT_FILE
Output file path
-m MODEL_FILE, --model_file MODEL_FILE
Model file path
-d DEVICE, --device DEVICE
'cpu', 'cuda', or 'mps' (default cuda with cpu fallback)
-b BLEND, --blend BLEND
Blend of input and output [0.0, 1.0] (defaults to 1.0)
-v, --verbose Verbose logging; multiple -v options increase verbosity to max of 2.
example usage:
uberSmooth.exe -m uberSmooth-dso-stars-v0.1.pth -i inputImage.tif -o smoothImage.png
There are similar quality tools on the commercial market, so I'm asking for a donation of *any* amount to help cover costs during development and model training. (Entering an amount under $1.00 USD will provide gratis download without any payments or accounts required to avoid processing fees). I plan to eventually release the model training source code and encourage community sharing of trained and fine-tuned models for everyone to use. If you would like custom trained models, hands on help processing your data, or access to training code now, please contact me via my email listed below to discuss possible arrangements. This uberSmooth inference/prediction CLI binary is released under the apache-2.0 license. The two model weight files included with the binary are released under the cdla-sharing-1.0 license, and are both fine-tunes of the original SCUNet paper model, upon which the uberSmooth v0.1 architecture is also based. All dependencies and libraries used in this software are subject to their respective licensing agreements and copyright notices as specified by their original authors. Special thanks and acknowledgments to NASA and STScI for providing high resolution image data used in training from their webbtelescope.org and hubblesite.org sites which may be freely used as in the public domain in accordance with NASA's contract. This version of uberSmooth currently works best on fully processed and stretched non-linear images. Simpy use your normal workflow but limit or entirely skip denoise and deconvolution steps. Run uberSmooth as the last step before publishing. It may work okay on other data as well. It is compiled with NVidia cuda support. My old NVidia 1070 GTX takes under ~2 minutes to process a full APS-C 25MP 6k x 4k resolution image. It may take over ~20 minutes to process if only CPU is available. Supported file formats include:
- TIF (even 32bit floating point RGB astrotiff - best quality available)
- PNG
- JPG
- WEBP
- possibly more...
- 32bit floating point FITS and PixInsight don't work yet...
- uberSmooth-planetary-v0.1.pth
- gaussian read/poisson shot/speckle/salt & pepper noise, gaussian blur
- sha1sum: 87ef0357f3ded5417ab23d31937f8b3c754c7091
- uberSmooth-dso-stars-v0.1.pth
- gaussian read/poisson shot/speckle/salt & pepper noise, gaussian blur, Kolmogorov atmospheric blur, slight optical aberrations (coma/spherical), star saturation and shape
- sha1sum: 4a1d8fa556b36927aa95d1b62e0a9f059c3c2df3
* email leimgrub ~at~ gmail ~dot~ com
* astrophotography astrobin
* coding projects github
* linkedin
* local astronomy club DVAA
* astrophotography astrobin
* coding projects github
* local astronomy club DVAA
Click image to briefly reveal uberSmooth image.
.
.