NoiseXterminator v3 is HERE and it's AMAZING! BUT... vs GraXpert, Linear vs Stretched data

[@CuivTheLazyGeek]

My Patreon: www.patreon.com/cuivlazygeek My Merch Store: cuiv.myspreadshop.com/ We have an explanation for the 12 minutes time of GXP! It seems that in PI, running GXP after NXT somehow makes GXP much slower. Running GXP without having used any of the NXT/BXT, etc tools is very fast (only a few seconds) and makes my GPU sound like a leaf blower :) Thanks to Andreas (www.tofahrn-foto.de/) for pointing it out to me! PixInsight: pixinsight.com/ NoiseXterminator: www.rc-astro.com/software/nxt/ Adam Block's video on NXT v3 (seriously, watch it!): kzbin.info/www/bejne/rGLZe3d8mNV2q9U For any affiliate link purchases, make sure to accept cookies if prompted! Amazon affiliate: amzn.to/49XTx01 Agena affiliate: bit.ly/3Om0hNG High Point Scientific affiliate: bit.ly/3lReu8R First Light Optics affiliate: tinyurl.com/yxd2jkr2 All-Star Telescope affiliate: bit.ly/3SCgVbV Astroshop eu Affiliate: tinyurl.com/2vafkax8

[@wattswheelhouse]

When I step back and think about it, I'm completely blown away at the amazing tools we have to capture and process this data these days. What a time to be alive

[@CuivTheLazyGeek]

It's mind-blowing right??

[@Sbach342Guy]

I just got into the hobby about 18 months ago and I can’t believe how much things have advanced in the short time I’ve been in the hobby. But I also can’t help but wonder sometimes how “real” my images are by the time I’m done with all the fancy editing tools.

[@wattswheelhouse]

@@Sbach342Guy I completely agree with your point in regards to the ai tools out there. But even manual editing tools are crazy powerful. The open source automation software is amazing. The (relatively) inexpensive hardware we have access to would have been unobtainable research grade instrumentation just a few years ago.

[@dumpydalekobservatory]

Interesting comparison Cuiv personally I don't mind a little bit of grain in an image, these Ai noise tools are really good but what I am seeing online are images that are so smooth looking they don't looks natural anymore to my eye at least.

[@CuivTheLazyGeek]

That's fair!

[@rrcroman]

Thanks for the detailed review, Cuiv! As another viewer mentioned, I'm thinking that the posterization (apparent low bit depth) ~22 minutes in are perhaps due to 24-bit lookup tables being turned off.

[@CuivTheLazyGeek]

I did try with 24-bit lookup after being told, but it looks like the issue is still there somehow, I'll experiment some more!

[@gazer63]

Thanks Cuiv! I’m looking forward to the upcoming updates to BXT and SXT.

[@CuivTheLazyGeek]

Me too!

[@danbrown3234]

A good HONEST review mate, well done! And thanks for covering the OSC data✌🏻I’m finding similar results 🫡

[@tolgagunes225]

I am huge fan of Graxpert , Seti Astro's Cosmic Clarity Suite and Siril. Competition will help us and helping already us ,the endusers of Astrophotography Softwares...

[@DarthPotatoes]

Seti Astro Suite is getting really good really really fast. I'm excited to se where it'll be in over the next months and years. Hope it'll give Pixinsight a run for it's money eventually

[@CuivTheLazyGeek]

Completely agree! Competition is awesome!

[@wigglybuff64]

Seti Astro is pretty awesome. When he released a video with new features I asked him about possibly adding a feature which siril can't do and within 30 mins he replied that he added it and for me try it out. Worked like a charm too and now I only use siril for stacking and use seti astro for everything else. The feature I asked about btw was to let you star align mono with osc frames for batch cropping. That way I can use osc data for stars only and mono narrowband for everything else.

[@dansstroshack]

Really good video (as always), Cuiv! I normally do Drizzle and Lanczos integration and then GraXpert and NoiseXTerminator before and after strateching (yes, loads of CPU/GPU work), and I found that GraXpert works best on linear data and NoiseXTerminator on stretched data (starless in particular). Neither can handle Lanzcos noise, which is a real shame. The slowness you experience is probably a bug in the RC Astro tools. They don't free up the GPU memory and forces all other GPU tools to use shared GPU memory (which is a incredibly slow (it's normal memory!)). I raised a bug report (94342) with RC Astro and it will be fixed in the next version. I just did a test on a 4072x2803 image and GX took 53 seconds, while NX took 18 seconds.

[@CuivTheLazyGeek]

You're right!! If I used GXP immediately after starting PI, it's super fast, and makes my GPU sound like a leaf blower!

[@sianikolaou5440]

That's very interesting, Cuiv, I now have to try both ways to see which does better at denoising with regards to starting with a linear image versus stretched. I trusted what the developers suggested with GraXpert!

[@valeryloriot2303]

I tested the v3 recently, was impressed by the denoising, especially on mottling ! It makes a huge gap for my data ! Now i have to try after stretch!

[@JonnyBravo0311]

At 22:53 did you try applying the 24 bit screen transfer? The rasterization you mention looks more like the screen transfer not handling it well instead of NXT producing something bizarre. I watched Adam's video right when it came out and was pretty amazed by the results. Russ took a single sub and after applying NXT, you would have thought it was hours of data stacked. I really like the ability to control the luminance vs chrominance noise as well as the frequency separation introduced with AI v3. With all the cloudy / rainy / snowy weather here, I've got plenty of time to go back to old data and play around :)

[@CuivTheLazyGeek]

I had saved the whole PI project so I reopened it and tried with the 24 bit screen transfer, but no change...

[@zer0g77]

The creator of NoiseXT clearly stated that the AI is trained on both linear and strached data. In his opinion you can run it in both stages with no preference. Basically it is stretching the data inside its process to remove the noise. Personal preferences may vary as well from image to image but in general no matematical reason to prefer linear vs stretched.

[@CuivTheLazyGeek]

Yeah if applied on linear data, NXT stretches the image, applies the process, then unstretches it. Russ says it should be fairly tolerant of either. From my tests it did seem like I would recommend using it on the stretched state

[@CuivTheLazyGeek]

I would say there is a mathematical reason though, especially if applied on each channel in the linear stage. Even otherwise, the stretch/unstretch within the process would require less aggressive mapping if the image is already stretched...

[@zer0g77]

@@CuivTheLazyGeek Anyway it seems to me more important to apply it to a RGB image to take advantage of the separate color and luminance settings and to use the full training set of the AI. By the way it is ony my opinion with absolutely no quantitative test. Just a guess...

[@quentinG_9841]

16:40 The preview has always been available with NXT, even with AI version 2, of course 😄

[@CuivTheLazyGeek]

Hahaha of course it has, I was just too dumb to recognize it lol thanks!

[@charliemiller3884]

$300 PI is not expensive. A $10,000 mount and $3,000 camera are expensive :)

[@CuivTheLazyGeek]

True!!

[@Undy1]

$382 PI (300€ + 23% VAT) is expensive when my mount with guiding is $543 and my camera is $160.

[@CuivTheLazyGeek]

@@Undy1 To be fair, thinking of prices in relative terms is a terrible habit that all sales and marketing tactics take advantage of. 300 euros is 300 euros, regardless of what mount or scope one has.... And it is indeed quite a bit of money. That said, if you ever have the budget, I would urge you to try the PixInsight trial just to get an idea of how good it is :)

[@Undy1]

​ @CuivTheLazyGeek Thinking in relative terms is good if you're comparing what value you can get at the same price. With my budget setup I'd be much better off spending $382 on an upgrade like a better camera (D5500 instead of D5100) or maybe getting a new filter (e.g. an additional SII or SII+OIII filter to supplement Ha+OIII which would let me do SHO palette) rather than PI which would only be a marginal improvement at best when compared to free software (and let's not forget that NoiseXterminator is a paid addon that costs extra too!). Getting the best value of your money is crucial. For someone with $10k gear PI might be the best upgrade they can get for $400 but not for me.

[@JamesBeckner-y7e]

Both NoiseXTerminator and GraXpert do a good job but by far the best noise remover that I've used is Photoshop's RAW denoiser. What it does that these other programs do NOT is that Photoshop doesn't fully smooth the background. The Photoshop AI seems to understand that you probably don't want a completely smooth or noise free image and what Photoshop does is reduce the apparent size of the irregularities and the noise-level color variations without any affect on the image sharpness or overall color in the image. So, you just get an image that has a much finer grained pattern instead of the blocky irregularities in the original. However, Photoshop's RAW denoiser only works on RAW images from Nikon, Canon, Sony and a few other camera brands and thus there is no way you can use it on calibrated or stacked images or even on still frames from astro-dedicated cameras. Too bad, since Photoshop's approach seems to produce much better looking results.

[@Alan-vk6bk]

I`ve been using using BlurX v3 and i`m over the moon with it!!

[@celiaxel3448]

Great video, thanks again! 👍

[@Apagadorable]

Great to see that this update is coming to the Photoshop version too. Of course Pixinsight is much better for astro images, but still.

[@CuivTheLazyGeek]

That's so nice! I didn't realize that was the case!

[@robkiefer822]

Previously, I prefer NXT. I tried Graxpert several times, however, I was getting too often too many "extra" stars. Plus Graxpert really takes long. NXT v3 - I first have to play around a bit, but the real strength of v3 will be in the control of luminance / chrominance noise as well as the frequency separation (all controlling colour mottling and details a bit better). Looking forward testing it with my images.

[@prookyon]

If Graxpert run on a medium size image takes 12 minutes then this for sure points to some issue on the system where it ran. Most likely the VRAM on GPU was full and Graxpert had to partially use system RAM. This makes it massively slower. Usually this can happen when some image/video edditing program has decided to grab all VRAM to itself. For example Affinity Photo does not release VRAM until you actually close it completely - just closing images is not enough. I have hit that isue several times since I am running by modern standards a GPU with quite meager VRAM (8GB). Photoshop, Lightroom, probably most video editing programs etc. all have the same issue to some extent - so good idea to completely close them before using Graxpert.

[@CuivTheLazyGeek]

I added to the pinned comment - seems a bug, it doesn't properly use the GPU if NXT or other AI tools have been run first. I tried GXP first thing, and then it took a few seconds (and made my GPU run like a leaf blower). The GPU in question has 12GB of VRAM

[@wigglybuff64]

@@CuivTheLazyGeek This is because when a model is queued it allocates all the gpu vram to it so even if it's idle it will still stay queued until you close it. This is common with ollama or studiolm or anything that loads ai models. Think of it as first come first served but won't release resources until shut down.

[@dadwhitsett]

Interesting. I just downloaded the update to noise exterminator yesterday but haven't had time to use it. The 12 minutes is a game changer for me.

[@CuivTheLazyGeek]

Apparently the 12minutes is a sort of bug on GXP side? Someone reached out to let me know, and it only happens after I've used another AI-powered process. I hope they fix it!

[@dwightlethbridge2796]

Fine review of these tools. I’d be also interested in your thoughts on SETI Astro Suite, another capable free product.

[@CuivTheLazyGeek]

I really need to test it out vs PixInsight :)

[@yearningmice]

Gotta update tonight and check it out!

[@CuivTheLazyGeek]

Hope it works great for you!

[@murraymacd]

Seriously impressive what can be done by these tools these days.

[@GilPaulbert]

Been waiting for your comparison/opinion on NoiseX update. Russell Crowman said, more updates are coming for BlurX and StarX soon aswell, exciting times !

[@CuivTheLazyGeek]

Very exciting times indeed!

[@mireia3208]

I actually like GX more. If you focus on the super dim stars, it seems that NXT is more aggressively removing them. Although to be honest, probably the way I will end up using them is to mix back like 10-20% of the original image into the noiseless image. I think it is not only me but a more general human perception thing that having just a tiny bit of random noise gives a sensation of having sharper (and better) image.

[@gregnicholl9011]

Using any of the AI de-noise tools on a linear image never works for me, it basically destroys the image. They have only ever worked for me after stretching. I must be doing something wrong. Thank you Cuiv, love your work!

[@CuivTheLazyGeek]

Interesting, and glad this was useful! Well, as long as it works :)

[@SDNoel73]

Would be interesting to see if you use NXT V2 first and then use V3 after what the outcome might be.

[@CuivTheLazyGeek]

Interesting idea 🤔

[@cyberfel2010]

Hi Cuiv, I recommend to rename the old IA model for your icons to take the new one by default

[@CuivTheLazyGeek]

Right, good point, thank you!

[@astrobondfr]

Awesome !!! :) ( topic and images ;) )

[@CuivTheLazyGeek]

Glad you liked it!

[@TheCongratulationsChannel]

I don't think GraXpert hallucinated the stars. I can see them in the original picture. If then your camera noise is already hallucinate stars. And is the Hubble image raw or already processed too?

[@l.m2517]

BlurX 3 that’s what I’m really waiting for!

[@CuivTheLazyGeek]

That's going to be amazing :)

[@bigguix]

lots of work ! merci ! but very useful information !

[@CuivTheLazyGeek]

Merci ;)

[@qo4kh]

Assume the next generation of de-noisers will reference Hubble-type photos to eliminate hallucinations etc. We now know what the star field should be making it easier to "clean the data." At what point will your image basically become the reference image?

[@anthonygondola3086]

I have been on the fence concerning PI for a long time. I am very comfortable with the usual free items for processing and I think I get great results. Still, I wonder if I am really holding myself back by not using PI. I'm sure there are a lot of people like me out there. I think it would be great if you could do a video like this one comparing the processing result using PI verses using Siril, Seti Astro, Graxpert, etc.

[@CuivTheLazyGeek]

To some extent I've done it with my Seestar and D2 processing videos - but a Seti Astro Suite vs PI is a great idea!

[@anthonygondola3086]

@@CuivTheLazyGeek Excellent and thank you in advance!

[@schweitl333]

personally i get the best results, when i stretch the greyscale images, then apply NXT and then combine the images. correct me if i'm wrong, but that is the only way you didn't try. or did you?

[@CuivTheLazyGeek]

Correct, I haven't tried that way! Something more to experiment with;

[@pompeymonkey3271]

In his interview with Adam Block, Russ Crowman did say to only use NXT on the individual channels in narrowband images. This, apparently, is because typically the signal strengths between each channel vary a lot.

[@CuivTheLazyGeek]

Yeah I recall that, but then experimentation (at least for me) says different, somehow... When combined we have indeed stretched each channel separately though

[@ninjanekonico]

12 minutes for graxpert denoise ? With my 3080 laptop gpu, not desktop, it takes around 30s, something may be wrong... I can't compare with the new nxt, but with setiastro, it takes around the time than graxpert

[@CuivTheLazyGeek]

Apparently, running an AI process before GXP denoise makes GXP denoise super slow...! Or at least so I'm being told. I'll need to test!

[@carvrodrigo]

How to make the image’s windows move together?

[@CuivTheLazyGeek]

On latest version of PI, hold Ctrl key when moving one image tab to another!

[@MrGeert1972]

I start with blurx correct only SETI Astro ADBE NoiseX SPCC and after all my other processing I run NoiseX again :-)

[@kurtkroneberger4540]

If NXT is better on the stretched images while GXP is better on the linear ones, why did you not compare these two?

[@deep_space_dave]

With the advancement in AI tools, it really blurs the line between astrophotography and art! Russ really outdid himself on this latest version! Thanks Cuiv!

[@CuivTheLazyGeek]

Absolutely!

[@andrewweller5119]

tbh i have always had better results using noiseXterminator on non-linear images rather than lenear. even with 24bit applied.

[@CuivTheLazyGeek]

Yeah, I tried with 24 bit STF but it didn't seem to change much

[@kowalskik.1333]

yeah

[@captainfruitbatify]

12 minutes for GraxPert? Something not right there. An 11 megapixel (IMX294) image takes 2 minutes on my PC, which is a 10 year-old Intel Core i7-6700, with a very modest nVidia card (GTX1650 4GB).

[@CuivTheLazyGeek]

Someone reproduced this, apparently, running an AI process before GXP denoise makes GXP denoise super slow...! Or at least so I'm being told. I'll need to test!

[@mgscheue]

I don't have Pixinsight (yet) but have been using GraXpert, Siril, and Photoshop. When the PS version is available, what would be the best workflow? Hit it with NoiseXterminator at the end? It looks like it works well with stretched images.

[@janossowski1490]

I'd say after stretch but not at the very end. You'll usually use color masks and selective color corrections in the process and you want the color noise to be gone before you do that because otherwise you can make it worse by these adjustments

[@mgscheue]

@ Thank you!

[@CuivTheLazyGeek]

Agree with @janossowski1490 :) right after stretch sounds like the best time to use it!

[@mgscheue]

@ Thank you! I might give the PS version a try when that’s available. Or download it and the PI trial and spend some time learning those.

[@MG-qj2zh]

Super comparison, but especially with Graxpert, the use of 3.0.2 is very problematic and it would probably have been better to use AI 2.0.0. and or cosmicclearity. I stopped fitting favorites for me because every tool reacts to the different sources

[@CuivTheLazyGeek]

You find it very problematic? I prefer 3.0.2 to 2 for my images...

[@MG-qj2zh]

​@@CuivTheLazyGeek Yes, because the 3.0.2 hallucinated with stars that are none and is very much -dependent.

[@CuivTheLazyGeek]

@@MG-qj2zh gotcha

[@MarcelBlattner]

But is it really desirable to have an absolutely noise free image?

[@Bills_APCh]

Many people seem to think so, and will pay hundreds, maybe thousands of dollars to get a zero noise image !!

[@EugeneRu777]

👍👍👍👍👍👍👍

[@TheCosmicGuy0111]

Niceee

[@CuivTheLazyGeek]

Indeed :)

[@antn8387]

GraXpert is better because it's free 99.

[@KZgun4hire]

I disagree comparing Graxpert to a star map and other images of that region. Garxpert has retained stars that Noise Exterminator has removed. Shouldn't assume one or the other is correct. Always compare to another good source.

[@CuivTheLazyGeek]

But that's what I did? I see stars that don't appear in reference images

[@keithburr3959]

I am sorry to say the number of adverts appearing in your videos are making them unwatchable.

[@457r0-q2y]

ublock origin my friend

[@Crimsonyte21]

Bruh, is there a free alternative for this? Like I'm not made of money wtf

[@marcinb493]

Graxpert is free alternative - i have both NXT and Graxpert and must say that Graxpert is amazing tool.

[@CuivTheLazyGeek]

Yes the one I spend a lot of the video time on: graxpert denoise