DeepSNR: New Free PI Plugin for Noise Reduction from the Creator of StarNet. You need this now!

[@setiv2]

You are going to want to max out youtube's playback quality for this one to see the differences!! Also if you are wondering why not increase the "Detail" slider on NoiseX. That is just an unsharp mask, it doesn't recover lost detail. I made a webpage to show that: www.setiastro.com/deepsnr-comparison

[@dominickzaucha]

Thanks for checking this one out! With regards to your pinwheel, do not do blurx/convolution before deepsnr, that's why you are getting those artifacts. Your results will be just as good as the others if you do deepsnr first. The software author suggests doing deepsnr as one of your first processing runs for best results.

[@setiv2]

@@dominickzaucha i dont think that was the issue with the pinwheel. I can do some more testing specifically on that one. Regarding blurx though you need to run that before any noise reduction. It is important since blurx is teasing out structure from the noise. If you get rid of the noise first bkurx cannot get its best results.

[@dominickzaucha]

@setiv2 I agree with that, except with deepsnr. It needs to be run early like step 2. I may have misinterpreted it along the way as well. It's just that I have the same exact noise pattern running blurx first.

[@setiv2]

I did the test. BlurX doesn't affect DeepSNR at all. I figured it wouldn't since BlurX doesn't mess with the noise (which is what makes it so amazing). You can see the comparison on my site: www.setiastro.com/deepsnr-comparison#deepsnrbxt

[@dominickzaucha]

@setiv2 thanks for the heads up! I'm going to have to do a few steps back and see why mine is doing the hallucinating.

[@dominickzaucha]

@@setiv2 Going back into further testing i think i found the culprit. DeepSNR only works well on 1x drizzled data in my case drop shrink 0.8 and Varshape 1.5 in WBPP. 2x/ or no drizzle etc all cause hallucination issues or poor noise reduction. The noise reduction on 1x data is stunning, I'm slowly reprocessig all the old stuff that I still have stacks of

[@Aerostar509]

I just tried DeepSNR on a simple RGB combo image and it completely removed the R,G and B hot pixels on a dithered, non calibrated channel combined image! Amazing.

[@robj144]

Wasn't this released a couple of years ago?

[@easyastroimages5818]

Damn. You don’t know until you know. The detail slider never seemed to do what it should on noise x

[@setiv2]

Detail slider on NoiseX is just an unsharp mask, it doesn't recover lost detail. I made a quick site so you can see the differences: www.setiastro.com/deepsnr-comparison

[@FrancoGrimoldi]

What about rising the "Detail" slider on NoiseX to try to match the detail retention while keeping the same level of noise reduction?

[@setiv2]

The detail slider is just an unsharp mask, it doesnt recover lost detail. I made a webpage to show what changing the "Detail" on NoiseX does www.setiastro.com/deepsnr-comparison

[@FrancoGrimoldi]

@@setiv2 Exactly what I was looking for, thanks for the effort, much appreciated!

[@mickwhitaker4086]

struggling to get this plug in on my Mac, done everything stated , but getting error in pixinsight .

[@setiv2]

There were a couple of options depending on what version of PI you have. Did you download the correct one to match your version of PI?

[@mickwhitaker4086]

@@setiv2I have the new version of pix. It’s showing up on the install module tab but when I press install it comes up with an error.

[@setiv2]

Did you grab the correct download? I also am running with the newest build of PI. What is the error you are getting? Did you get the one with the tensor flow file or not?

[@mickwhitaker4086]

@@setiv2 with the tensor flow files. *** PixInsight API Error: Module: /Applications/PixInsight/bin/DeepSNR-pm.dylib: Module initialization error: Exit code:1 ** Warning: No modules were installed.

[@mickwhitaker4086]

Yes with the tensor flow files. The error message is as follows. Installing 1 module(s): /Applications/PixInsight/bin/DeepSNR-pxm. dylib < failed *> PixInsight API Error: Module: /Applications/PixInsight/bi0/Deeps.R-pxm. dylib: Module initialization error: Exit code: 1 ** Warning: No modules were installed.

[@tomhoskins4913]

I followed the instructions exactly, it installed perfectly and I have the latest version of Pixinsight. The process showed up . I ran it on an image and the it executed perfectly according to the process console. One problem, it did notion to the image, it remained the same with no change at all. I did have an existing file for tensor flow in my bin folder already. I even swiped it out for the one DeepSNR provided. No help I reverted back to my existing tensorflow

[@setiv2]

Interesting. Could you share the image it wasnt working on? [e-mail]

[@Zegery]

must be run on 16 bit images or it will not work. save it as 16 and it will work

[@AstroIsland]

@tomhoskins4913 I had the same issue. Try splitting the channels, run linear fit, combine the channels back to RGB then run Deep SNR.

[@qx3V45p]

Works amazingly well and leaves all that small detail! Some reason it's very slow on my rig- 2 minutes to denoise an image. Graxpert DeNoise and NoiseX only take seconds. Not sure the issue. Hopefully I can get it to run faster!

[@setiv2]

What kind of computer do you have? Maybe on a mac it doesnt use the unified gpu acceleration.

[@qx3V45p]

@@setiv2 Windows 10- Ryzen 5 5600X and Vega 64 gpu. Kinda old but still works very well . StarX utilizes GPU very well with an older version on Photo Shop. Newer version installed on PI doesn't :/.

[@qx3V45p]

@@setiv2 The CPU supports AVX.AVX2 instructions

[@setiv2]

@@qx3V45p can your gpu enable cuda acceleration? That is the only way I know of getting PI processes to be GPU accelerated. If blrux and noisex in PI are accelerated so should deepsnr. I am not familiar with Vega GPU

[@qx3V45p]

@@setiv2 It's an AMD GPU so no CUDA. NP my friend, I'll look into it. Thanks for everything and all your hard work creating your great scripts and vids!

[@walterhock7809]

It's a fantastic tool, definitely! But there is a substantial speed problem with my (intel i7, 2021) iMac (Sonoma 14.5). NoiseX takes 12sec, DeepSNR 8min(!) Installation was no problem, copied the two files into /bin and the recommended other 2 into /framework. Is there any remedy to speed things up?

[@setiv2]

On a mac noisex uses your unified memory and gpu, sounds like deepsnr is just using your cpu. I dont have a mac, so for me it uses my cuda acceleration. You may want to reach out to the developer and see if there is a remedy for gpu acceleration or if there is a install step missed.

[@walterhock7809]

@@setiv2 I'd like to. But can't find the developers address.

[@setiv2]

@@walterhock7809 go to their website, they have a contact me page.www.mikitamisiuraphotography.com/contact

[@biodan8577]

On a M1 Max running Sonoma, it does not install- will send this error to the developer > PixInsight Core 1.8.9-3 Ripley (x64) (build 1612 | 2024-06-25) Installing 1 module(s): /Applications/PixInsight/bin/DeepSNR-pxm.dylib *** PixInsight API Error: Module: /Applications/PixInsight/bin/DeepSNR-pxm.dylib: Module initialization error: Exit code: 1 ** Warning: No modules were installed.

[@walterhock7809]

@@biodan8577 I'm not an expert, but there indeed may be an issue with DeepSNR-pxm.dylib. There is a question mark in the process control aside this module, all others have a green star. Again: It works, but extremely slow.

[@baa-austria]

Tried to install on Mac Mini M2. put the desired files into bin. The libtensor* into framework (AFTER ALLOWING execution!). Then executed ADBE and got mallocation error... Trying to use the alias pointing to the newer libtensor* - this does not work. So - the installation process of this tool will have potential and quality check also using the more recent lebtensor* from PI... Anyway - referring to the absolute ingenious ADBE some of us would be VERY happy since for example EZ denoise would get some "service" since it works without AI... (it is just a great dream ;-) )

[@setiv2]

I am not very familiar with mac's. Since StarNet and tensorflow libraries are native in PI you should be following the process for "Already Have StarNet". Maybe that will resolve your issue. Otherwise you can reach out to the developer. I never used the EZ suite. Is the repository still up? If he isn't updating those scripts anymore maybe I can take a look at the EZ denoise.

[@jimcollins1922]

According to the readme text with it, there is a dependency on AVX and it appears that AVX is not supported on the Apple silicon. I tried installing it on my M3 and it would not install. Too bad.

[@setiv2]

@@jimcollins1922 maybe time for an upgrade! PI v1.9 will require AVX2

[@jimcollins1922]

@@setiv2 Looks like operator error on my part. I downloaded the wrong version. FYI The apple silicon macs are the latest macs and they don't support AVX but pixinsight uses apples graphic service which I believe is called Metal.

[@biodan8577]

On my M1 Max, while DeepSNR would not install, the libtensorflow* libs did not affect the ability of ADBE nor BlurX/NoiseX to run

[@jimcollins1922]

Looks great and I tried to install it on my M3 MacBook pro and it will not install because the M3 does not support AVX. To bad because it looks great!

[@setiv2]

Time for an upgrade! PI v1.9 will require AVX2

[@jimcollins1922]

@@setiv2 Doesn't sound like a good economic decision if pixinsight won't support Apple PCs. I doubt that pixinsight will not have a solution to continue supporting apple. If they don't there will be a lot of ticked off users.

[@rocketcityastro]

If it's free it's for me. Thanks for another great video.

[@bezain7663]

Interesting I gave it a try . Smoothing is not so waxy compared to NXT, and that's a good thing. Unfortunately, it gave me some weird artifacts, like pure red or yellow hit pixels in the core of saturated stars

[@setiv2]

Do you own blurx? If so you can run that on correcr only see if that fixes those pixels. Other option if it seems really bad 8s remove the stars first prior to running deepsnr

[@AstroIsland]

Wow! RC has some stiff competition now. All that’s missing now is a BlurX alternative 😬

[@setiv2]

Russ needs to get on that new noise AI weighting update!!

[@JLCubing]

for deepsnr to work good u need to properly 1x drizzle so thats why it prob didnt do good

[@setiv2]

I always drizzle, not sure what you mean didnt do good, it beat noisex for almost all rgb images

[@JLCubing]

@@setiv2 I think I have bad eyes then my bad, but for me I still get some weird artifacts after using deepsnr

[@AmatureAstronomer]

Interesting.

[@captaincook6666]

Wish I'd not paid for NXT!

[@setiv2]

You still need it for greyscale and off the wall narrowband, but RC needs to step their game up quickly!!

[@captaincook6666]

@@setiv2 also I like the linear option...I read that NXT is meant to be for non linear is that true?

[@setiv2]

You heard very wrong. NXT is for linear data. It was even trained on linear data in the ai2. Ai1 was trained on nonlinear, but it did a 64bit internal stretch then denoise then stretch back to linear. Either way NXT is supposed to be used with linear data

[@setiv2]

All of RC astro tools work best on linear data, blurx, noisex, and starx

[@captaincook6666]

@@setiv2 ok I use bxt on linear. I don't have StarXT. I've been getting good results with NXT in any case...

[@nikaxstrophotography]

Well the installation process is what kills it for me, there must be a simpler way than this.

[@setiv2]

This is how you install StarNet++ and a few others. That is why I covered it if you werent familiar with installing a standalone process. Since I am not the author I cant add it to my repository. you can reach out to the creator if you are hoping they can do a repository or something.

[@vastrop345]

You can use RCAstro Cuda/Tensorflow GPU acceleration repository (as detailed on the Pixinsight forum). That should set up your system, assuming you have a supported NVIDIA GPU, for GPU acceleration. Once this step is complete, copying two files (the dll and the weight files) to the Pixinsight bin directory and importing the process is all you need to do. Feedback about that method seems positive on the Pixinsight forum. I haven't tested it directly myself as my system is already set up for multiple CUDA versions.

[@setiv2]

You dont need to do the cuda set up. PI has a native cpu tensor flow dll. You really have to just drag and drop those two files. If you want to do cuda setup then yes Russ has a repository for that portion, but technically it isnt needes to run this.

[@vastrop345]

@@setiv2 yeah, I know that but couldn't imagine someone finding the copying of two files being a killer. That being said, IMHO; you really want GPU acceleration, especially if ones want to compare results and experiment with different settings.

[@setiv2]

One persons simplicity may be anothers complexity... 100% agree if your system supports gpu acceleration you need to unlock that power!!