GraXpert Denoise AI 3.0 - not your everyday tool.....

[@steffen1181]

Hi Sascha, as already pointed out, v3.0 is a state-of-the-art architecture which was just published last month in a new publication. This version is meant for people who have a good GPU or do not care about computation time. Also, it was trained on OSC images only and we have already seen that it does not perform well at all on other images such as dual narrowband or mono. The framework we are using to run the AI supports less operations on macOS than on Windows. Therefore, on macOS it runs much slower. We can't really do much about that. Apple's CoreML framework for acceleration caters to a much smaller audience than for example Window's DirectML framework, and is therefore less supported by the whole eco system of machine learning frameworks. Especially the new v3.0 architecture uses some operations that are very new. Even though this version is slower and only performs well on RGB images, we wanted to release it for those who like to experiment and have fun tinkering with new tools. Also, it is very useful to get feedback very early in the development of a new AI from a broader audience which is why we released it.

[@viewintospace]

Thanks Steffen for your input!!! First of all, that it only works with OSC RGB data is an extremely vital information which Frank did not include in his message. I will add this now in the description in the hope people will read it. But then - about the whole Mac issue - it for sure can be done as Russ perfectly adapted his tools for MacOS. And given I assume this new framework seems to be the way you want to go and this is kind of a pilot, it will be key that you can also make it run on Macs decently, as 45 mins is just an absolutely impossible number given how fast it runs on PCs.

[@David-PB]

​@@viewintospace Hi Sascha, I cannot follow you here: As I have explained to you on the Dark Matters Discord 1) for platform-independent model inference and packaging we rely on onnxruntime since otherwise the development effort would be somewhat unmanagable for a hobby project like GraXpert 2) onnxruntime does not (yet) support all required DNN operations on CoreML (=macOS) as it does on DirectML (=Windows) 3) working around this takes (our free) time. Long story short (again): we are working on the issue, however, finding a workaround does take time (and a bit of patience while waiting for new onnxruntime releases). CS, David

[@viewintospace]

Hi David First of all I would not have released this AI without a HUGE disclaimer. It it practically not usable on Macs and also works only on OSC RGB pics, which make up for about 20% of all pics.... So given people usually simply upgrade when a new version is released I needed to create this video to make people aware of it - so in other words: I did your job... 😉

[@David-PB]

​@@viewintospace Hi Sascha, overall I tend to disagree, let me explain: it is true that our documentation on new ai model releases is still lacking. This gap is being worked on: Upcoming releases of GraXpert will contain a model manager to manage locally installed models (i.e. downloading, reviewing, uninstalling, docs). In the end it's a question of available capacity when that feature arrives (i.e., free time - I begin to repeat myself...). However, on your "need" to create your video: what would have been helpful for the community would be spreading the word on 1) the reason for the current macOS performance 2) the fact that it is a known issue to us and beeing worked on (both of which is information you did have beforehand). On our "job": we take the role of developers in an open source community and aim to invest our time in the best way possible for all those hobby astrophotographers out there - but yes, for sure there is room for improvement (e.g., writing that novel on onnxruntime feels like a waste of time by now). I for myself will do exactly that and get back to "work" now. CS, David

[@astrofromhome]

@@viewintospace Sascha, I am totally with you what you write. 👍Under my old name on YT I often disagreed with you what led to very good discussions. But here I share your opinion 110%. The communication from the GraXpert must have been better. Denoising is often a trial and error activity to see how much of it still looks natural and preserves the details of the structures. Such is not really possible in a reasonanble time frame with GraXpert Denoise AI3. I upgraded my PC to not wait for several minutes for any of the processing steps to complete. So this AI3 is absolutely nothing I am going to use as I will not go back to 5, 10 or even more minutes to execute one processing step. I stay with NoiseXTerminator as it gives brilliant results within 15 seconds for me on photos with my IMX571 cameras. As long as not other denoising tool is capable of delivering the same speed with the same or better result there is no reason for me to change. What I see as part of processing time to denoise is manual preparation time. So that I have to save the picture, load it in another applicaction, save from there and open it in PI to continue. So the tool or process needs to be executable in PI (what is the case with GraXpert Denoise) to at least have the chance to be able compete with NXT. But generally speaking the dedication of the GraXpert team is amazing!

[@stefanschneider3681]

MacBook Air M1: 2.0.0 two minutes thirty, , 3.0.2 two HOURS - but I was running starnet ++ in the background for part of it ... Thx for the video! And I think the Graxpert-people get you wrong! We are all extremely excited about the tool, it being free, them putting all these efforts into it. But just be honest about it! I watched one video of the developers and their "Now this takes a little moment" - cut - "let's go on with show" made me chuckle 😅

[@ctortolini]

Hi... On my Windows laptop, NoiseXTerminator takes 5 seconds to denoise a full 7200x5500 pixels image. GraXpert Denoise took 95 seconds. However, I must say I like the result much better. NoiseXTerminator tends to make stars "puffy" and also completely erases the fainter smaller stars. GraXpert managed to keep the stars sharp and kept the fainter/smaller stars visible. I think a minute and a half is completely worth it.

[@amp2amp800]

Hi Sascha. I've been using it for a few weeks now. The hardest task I could give it (my x2 drizzled 2600mc pro image of tadpoles and flaming star - 6248x4176 ) took 10 minutes for a background extraction denoise and save on a Windows 11 PC, when run via the GraXpert UI (I normally just fire it up inside PI). I upgraded my PC a few months ago to get the power it needs for PI's AI tools - but on a budget, so my PC has close to the absolute minimum capabilities - 11th Gen Intel(R) Core(TM) i7-11700 @ 2.50GHz 2.50 GHz 16GB RAM - NVIDIA GeForce GTX1050Ti. CUDA is installed and enabled. I honestly didn't notice any real issue with undrizzled images or 3008x8008 images from my 533mc pro, and I've grown accustomed to spending a couple of hours to get raw data to a first draft 'final' result. This isn't killing me, but it IS discouraging me from experimenting with non-default settings.

[@viewintospace]

I believe if you use it "for weeks" now then it is not AI 3.0 as this one was only released these days.

[@amp2amp800]

@@viewintospace Yes you are very likely correct.

[@alandyer910]

Thanks for your report. On my M1 Max MacBook Pro GraXpert took about 2 to 2.5 minutes on 30 to 45 megapixel 16 bit TIFFs. Longer than any other noise reduction software by a factor of 5 to 10, but not unacceptable. Certainly not 45 minutes! But it can only import and work on one image at a time. You can’t batch process a set of images - I was trying to process a test set for a review of NR software. Nope! Can’t do it. One at a time only, if you please! So that slows things down more.

[@elythomaslumber]

...used an i9 Mini-PC / 32GB RAM without a graphics-card (CPU-internal GPU only). It needed 73min for 26MPixel 2 x Drizzle picture) 🙂 Just for information! (Default configuration!)

[@meibergstrmandersen9181]

Thank you for reviewing it. I clocked 7 minutes and 40 seconds on a Macbook air M2 processer, and a OSC stacked in ASI. It is indeed slow. Better quality yes.

[@viewintospace]

Denoising with AI 3.0.0??? - no way in 7.4 secs - not even on a PC....

[@meibergstrmandersen9181]

@@viewintospace i meant to say 7 minutes and 40 seconds - corrected in the comment 😊

[@Joel.DeepSkyForge]

Sascha, from PixInsight, when you re-execute GraXpert Proccess (not the Toolbox Script) on same image with a different strength, the result should be immediat, like in application. If that not the case, please let me know.

[@darkoskrlec3101]

System: Dual Xeon 2680 V4, 128 GB RAM, Quadro M6000 24GB Linux: 54 min Windows 10: 11 min Sample image: 10865x7236 px RGB

[@HaraldKrause]

Thanks again for your video. 3 minutes for a 6000x4000px image. Tensorflow, Cuda etc. installed (via pixinsight repository from rc). Lenovo P16 with 12th Gen Intel(R) Core(TM) i7-12800HX 2.00 GHz. 64 GB Ram

[@nikaxstrophotography]

I'll stick with Graxpert for most gradient extraction but definitely noisexterminator for noise. I get excellent results and if I want even better noise reduction and decent images then in Photoshop Topaz denoise used sparingly

[@shubinternet]

I’ve heard that NoiseXterminator runs much faster on M* series chips running MacOS, due to the much more efficient library that they are using - as compared to PCs.

[@viewintospace]

That is accurate - the RC Astro Tools make full use of the M-Chip architecture and are running blazing fast - as does AstroPixelProcessor (APP)

[@KJRitch]

I have a M3 MacBook Pro 14" with 36g ram , 1tb SSD with MacOS 14.4. I've been playing around with GraXpert for a little while and just upgraded to their AI version. I didn't realize it was version 3.0.2. Is that your version? My camera is an ASI071MC Pro, only 16mp. I tried the noise reduction on a stack of 89mb, (60 frames of 180sec exposures of Leo Triplet) . I performed a background extraction AI at 0.5 level and denoise at 0.75. It took about 1 minute and 18 seconds. Maybe I don't have the latest version?I notice the the file started at 89mb and now is 175mb after GraXpert back ground extraction and denoise. I wonder why the file size grows so much?

[@viewintospace]

Version 2 is correct. In Advanced Settings to the right you can change the AI model to 3.0.0 and you will see how much slower it gets. No clue why it grows so much in size....

[@KJRitch]

Didn't realize it was in the advanced settings. Just trying to download the AI model is very slow , such that I'm getting application not responding. Guess I'll wait for the next release. Thanks

[@KJRitch]

I tried GraXpert again this morning. I noticed there is a new AI denoise version 3.0.1. I ran my 89mb stacked image through GrapXpert gradient and then Denose with 3.0.1 AI set in Advanced. It prompted me to download and it wasn't struck like before. I'm on Starling so maybe less traffic. Once the download finished Denoise started. I estimated it took between 9 and 10 minutes to process my image, Leo Triplet. Three galaxies, a lot of background and not a large field of starts compared to nebula. Maybe those take longer. Again, M3 MacBook Pro, 36g ram

[@iAribeth]

Tried it out on my PC, took a few minutes to process a 5000*4000 pixel image. I have a RTX 2070 GPU and hardware acceleration is enabled.

[@spacepunch72]

Cool! So my WIndows 11 PC does the job in exactly 1 minute with AI 3.0. Image from Altair Hypercam 269 TEC OSC 21mp camera. It's a 16 core AMD Ryzen 5950 with 128GB RAM and a 12GB RAM GeForce 3080 GPU. And that is the part that does all the work, CPU is bored and the GPU is at about 60-65% load. Denoise was set to AI3.0 and level 0.84. Also background extraction is very fast, just a few seconds. So seems the Mac Metall is not supported here and all is done by your M2 CPU, which obviously is a lot slower. Hope this helps.

[@viewintospace]

That is quite a machine that you have!!!!!! Any yes, that is exactly the reason and I guess they can not even really activate the multi-core as my CPU was only on 50% when it was running....

[@shubinternet]

I’d love to see a head-to-head comparison of both tools on macOS versus a PC, on small, medium, and large images, both for OSC and monochrome images.

[@viewintospace]

Quite honestly, it would be meaningless, because what I can see now based on all the comments is that on the PC side the durations differ massively, mostly based on the Graphics cards. And same would be with NXT - CUDA enabled or not makes all the difference on the PC side.

[@danielpetzen]

Thanks for another excellent video, Sascha! I'm not sure if I'm doing anything wrong, but GraXpert is faster than NoiseXTerminator (when denoising at 0.75%). It takes less than 2s for a 4000x2652 OSC image. The code signing is invalid and I've not had the time to work around that, so I had to run it stand-alone, which meant I couldn't get exact timings. It's so fast that the processing time isn't noticable. The results are really good, so I'm looking forward to explore how it stacks up against NoiseXTerminator. I have an NVIDIA GeForce 3090 (it's the speedy 24GB version, though). My CPU is a 4GHz AMD Ryzen 7 (and 64GB of memory), but the CPU is rarely a factor nowadays.

[@viewintospace]

Are you sure you are on AI 3.0.0 ???

[@danielpetzen]

@@viewintospace I set Denoising AI-Model to 3.0.0 in the advanced settings. AI Hardware Acceleration was already defaulting to "Enable acceleration". I just ran it again after checking the settings, and it crashed. I restarted it, checked the settings and tried again and now it took over two minutes. I lowered the denoising value to 0.65 to see if it was some threshold, and it now took less than 2 seconds again. I went back to 0.75 and it still took less than 2s. I reloaded the image and started from scratch and I was back at 2m20s time to denoise (0.75) at 100% GPU and full VMEM utilisation! I'm completely confused here. Are all subsequent denoising re-using the initial denoising? Even if that is the case, then how could I not have noticed the first denoising when I tried it the first time? I'll have to look into this later today. I'm sorry if I caused any confusion around performance. I think it's safe to say that denoising a 4000x2652 OSC image on a NVIDIA RTX3090 seems to take a whopping 2m20s for now.

[@danielpetzen]

Correction! I found the problem. Pixinsight was running in the background and had gobbled up all VMEM! GraXpert was running in shared GPU memory. Denoising takes 25 seconds on my hardware (that's why I didn't notice it when I first installed with everything else thatw going on). Still more than 2 seconds, though... :-(

[@danielpetzen]

...and it drops to less than 20s if I increase the "AI inference batch size" to 32. It uses more VMEM (of course) and utilises the GPU slightly better (70%->78%).

[@robertgrenader858]

Is there a link to the repository for running Graxpert as a script in Pixinsight? Or does it install from the download?

[@viewintospace]

Link to add into the PixInsight Repository: pixinsight.deepskyforge.com/update/graxpert-process/

[@robertgrenader858]

@@viewintospace Thank you. After going through the convoluted Pixinsight upgrade process, I run the Graxpert DeNoise process, and I get an error report that terminates the process. Background extraction runs without issue. I will check with the Developer

[@vastrop345]

@@robertgrenader858 you need to run GraXpert standalone at least one, go to the advanced tab on the right and do a denoise with the AI model you want, the model will be downloaded at this point by GraXpert and then will be available from the Pixinsight script for future use. Even with GraXpert installed and the script installed in Pix, the model will not be downloaded automatically.

[@Joel.DeepSkyForge]

Point closed after upgarde to GraXpert App version 3.0.2 which is the minimal version for denoising AI 2.0.0 or 3.0.0 ;-)

[@JeffHorne]

I wish you would have used a noisy image to begin with. I can’t really tell a difference between before/after.

[@viewintospace]

The difference between the difference denoising option you would have anyway not seen. The KZbin compression is really a huge drawback for such videos.... 😔

[@KosmicznyTom]

amd 3600, asi2600 5926x3639 64gb ram, but shared about 4% basic and 11gb ram the process takes 2 minutes

[@janeclark1881]

Interesting. I use OSC and process on a powerful Windows gaming PC. So it should work for me. I wonder about wideband IR shots? I sometimes photograph objects using 750-1000 micron IR.

[@janeclark1881]

I spent my working life as a mathematical modeller living on the droppings of the PC gaming industry. Now it's that way with my astronomy. Funnily enough, I can't stand computer games.

[@viewintospace]

@janeclark1881 You are the person I waited for!!!!!! I have the Antlia R+ filter, which in principle is exactly that - a near IR filter cutting off below 750. But for the life of it I could not find any meaningful way to use it until now. The only thing I know of these filters are good for is planetary and moon shots. So if you have any deep space use cases for this filter, could you PLEASE tell me????

[@janeclark1881]

@@viewintospace I have done some deep sky IR photography. I have a lot of ideas on the subject. The "interesting" slightly tricky part is that some of my images are in a book manuscript I have just completed; and I am submitting it to a publisher per a contract I have with them. I clearly have a political need to keep them onside. I don't think it's that much of a legal issue. I wonder if there's a way we could have a very brief offline conversation about it?

[@viewintospace]

@@janeclark1881 Sure, happy to! My email is [e-mail]

[@Excessus1978]

Its quite fast on my PC and I like the result

[@vastrop345]

a bit under a minute on a 24MP image (2600MC) on a now relatively old Ryzen 3900x but a relatively new NVidia 4080. That being said, I don't like the AI 3 result that much.

[@jamespeirce2582]

Why would they release a program this bad? With such horrible optimization (especially for the really excellent Neural Engine on macOS)… with terrible and excessive noise reduction (I cannot think of a single reason why anyone would ever want to remove all noise, or so much noise)… this needed more time to bake in the oven.

[@viewintospace]

I think we need to consider a) that the Software is free b) that they develop in the cutting edge tech area c) that they are no marketing/comms pros Bottom line is that for Windows users with high powered graphics cards we want to denoise OSC RGB pics this solution seem to be working quite well. But they should have made it much more clear with the release, that this new AI is really only for this small group presently.

[@jamespeirce2582]

If the noise reduction works as aggressively as in the video it’s… just bad. “Free” still ought to be open to critique for validity as a tool to be used.

[@parkofile1173]

I am sure I have a much smaller image than you are using but with AI Hardware Acceleration set to "Enable Acceleration" w/Apple M1 Max 32GB Sonoma 14.5 Reading image: /var/folders/cb/3tyntb295914n641z4gccgkm0000gn/T/M101_Stacked_07194_clone_GraXpertImage_Temp_GraXpert.xisf Loading image: w=1080 h=1920 n=3 RGB Float32 PixelMath: Processing view: M101_Stacked_07194_clone Executing PixelMath expression: M101_Stacked_07194_clone_GraXpertImage_Temp_GraXpert combined RGB/K channels: done Truncating to [0,1]: done 100.049 ms. Are my results with "Strength" set to 1.000 with V3.0.0 Batch Size 4 I am missing a setting? Thank you