วีดีโอ

Interesting! So, if you stack images, do you just average pixel by pixel? If you have a plane /bird/ satellite in one images, it can destroy the whole result. How about using a median and not the average?

Great video! I want to see more! From the signal theory perspective, averaging the pixels of 3 images reduces the noise to 1/Sqrt(3)=0.57 or 57%. This is better than taking just one random image. Taking the median and not the mean reduces the noise to only 70% (I made the calculation for gaussian noise). So averaging looks better. However, this is only true for gaussian noise. You can have a case where the noise affects only one image. Taking the median pixel by pixel will eliminate it. Lets say the birghtness values are: 50, 53 and 89. Averaging gives you a brightness of (50+53+89)=64. If you take the median, you get 53, since its the number in the middle. Do you have an option to take the median and not the average of the photos? Of course, you should take the median over much more than 3 images :) Thanks a lot!

i like Your videos - keep up the good work man !!

John Hayes, an optical engineer says 4.8u pixels would be the perfect pixel size. No such camera like that exists.

I have a error when i wanne run the app it says error package or namespace load failed for ‘imager’

Thanks, this is great!

Stream starts at 4:10

I just watched this video a couple times and I still think that there's something missing. Let me give you an example let's say you have an F5 40-in telescope, and an F5 80-in telescope. In this equation you're suggesting that they both get the same amount of signal. Because they have the same focal ratio, but what about the extra light the 80 mm is collecting. Are you saying that that light is just lost because of additional focal length?

starnet++ is not allowed, right? are there any other non air based similar tools? :(

Interesting challenge. Please however, stop editing your videos with the ultra rapid cuts from one sentence to another. It makes it extremely unpleasant to listen too. It makes me nervous to listen to your dialogue that I have to stop watching mid video. It’s a shame because your content is excellent. Thanks

Almost missed this one. I feel partially responsible for this one😂

Awesome video! It was really fun to work with the data. I've never used the lens distortion feature on my telescopes, I usually just just it on my photos from a camera lens. Thanks for the idea! I have one scope where the flattener isn't perfect so I need to try it out.

Great walkthrough Mark!! So many new ideas from your workflow!! 12:45 there are no coincidences 😂 it was the DSA t-shirt! Cheers! Doug

Hi Mark. I spotted an error in your pixelmath expressions. In step 3 the blue should be B+... Rather than R+... I really appreciate your covering of Siril and Gimp. Useful stuff like correcting for flatness that's new to me.

Excellent walk through Mark! Loving your new wardrobe choices too! 😜

Using dust and scratches filter in Photoshop can act as an starnet and bring the galaxy much better than your edit.I'm still impressed by the Human's ability now.

Wow, probably the most informative processing rundown I've seen in a while, thank you so much! I practically excusively use Siril+Gimp (with starnet & now GraXpert) and there was so much new information in this video that I'll probably go and reprocess a bunch of images (and/or try the challenge!). If you had to guess, what did you do differently to bring out the star colours so much more nicely in the second version of the image?

Error in the video at 6:40. The expression for the B channel should be B + Q*(HaR-med(HaR))*.2, not R + Q*(HaR-med(HaR))*.2 Also note that there are definitely things I could improve on in the video, like doing deconvolution on the LHa image instead of just the L. Or not removing the green noise until I've cropped things. But I hope this video gives you some ideas for how to process the data so you can improve on what I've done! You've got real intelligence, after all! Good luck on the challenge :) The Pixelmath Expressions: Step one: Create Ha with R Subtraction Layer Single Expression: Ha-Q*(R-med(R)) Q = A number you choose. Step 2: Load the new HaC layer into Pixel Math Step 3: Create the Ha RGB layer R + Q*(HaR-med(HaR)) G B + Q*(HaR-med(HaR))*.2 Q = A number you choose. Step 4: Create LHa layer L+Q*(HaR-med(HaR)) Q = A number you choose. You can support me and the channel on buymeacoffee.com/deepskydetail :)

Hola, podrías comentar que Sistema Operativo le instalas a la RBPi5? AstroArch? Astroberry? muchas gracias

Great challenge Mark! One of the reasons I stick with Siril is that it pushes me to the basics of processing, understanding what is at it's core. What is it that I'm doing? I have used the AI tools a lot too, lately. Mainly I think because they save so much time! However... looking back at older work, mind you I started this hobby just a year ago, I started missing the fluffyness appeal of more less perfect things, kind of a sterility that creeped in some of my pictures. Specially stars I guess. And I want to get that back. So yes I absolutely get what you say. AI is stunning in some ways but not everything. Personally I think SetiAstro's statistical stretch is awesome, not AI as far as I know, and one of the great tools to help consistently stretch an image. He's a powerhouse, great work , can't keep up. 😅 Working with the star mask in your program, I think I never really understood how to do that, and the more I play with starmask the more I think that being able to adjust it properly is important to the end result. Maybe you could do a revisit on it ? Really great work. Keep it up! I learn so much from you.

Depends on different factors and knowing your imaging train, and what you’re imaging. There’s really no one size fits all.

How about everyone submits 2 images? One with AI, one without. Definitely a fun challenge, but I doubt anyone is going back to older processing techniques if the newer ones make better images...AI or not!

Hopefully the data was from Bortle 9 to make it a real challenge!

Nice video again! I think I read many misunderstanding concerning 32 bits float vs 16 bits in comments. People are talking about better "distinguishing" between, for exemple, 5.5 and 6 value in 32 bits. In digital domain it's not a "precision" matter concerning data between 2 different values, BUT a matter of increasing dynamic range for a better calculation in lower values (not in between them), for them not being lost, making increase levels possible afterwards, which is the case by stacking and stretching in processing. What matters is the ability for the sensor to cover enough dynamic range and to have the sensibility needed to catch the fainter photons, and then in processing to get datas contained into 16 bits range at the output at the end of the process, which is more than enough considering our eyes can barely see 10 bits of dynamic range equivalent. That's why all monitors are max displaying 12 bits (in Dolby Vison for instance).

Great idea, I’m a OSC guy, but this was fun to play with. Just curious, what was the camera? The pixel size is huge, 9 microns, or was this on Bin 2? And the exposure times? The luminance image is a bit overexposed, there’s lots of saturated pixels. Thanks again.

not using starnet!? that is going to be hard lol

It's sort of interesting. ("Hey ChatGPT, what are some non-AI algorithms that can clarify my telescope data?" JK 😅) We're definitely already heavily reliant on machine learning / neural networks in every aspect of tech. It's just so easy.

it's high time to ban the use of AI in astroprocessing, after these tools a photography becomes a drawn picture great competition! but who needs a Bakhtinov mask? all astrophotographers use autofocusers

Ugh... manual deconvolution... how 20th century :P

Thank you for exposing this reality, last year when i started imaging i quickly realized that most images shown where literally AI corrected to the point where people seems to do d..k measuring... Nice challenge!

Nice and valid points, you should definetly have more data in same conditions for better comparision and understanding, maybe in next video ? If you are interested I can provide some data for Nikon Z6 and IMX 585 Cooled camera.

Challenge accepted sir! :-)

This sounds fun but feel I don't know enough yet to do this. Plus I have to figure out what I can't use.

I did manual deconvolution back in the day with a underpowered pc not again. it took me 4 days with crashes and other things. Not even good price can not make me do it again even with a good pc. I am fine with everthing else but deconvolution bane of my existance.

Great challenge! Looking forward to trying out your data.

Old school

Nice! One aspect that gives me quiet pause is the HA composition. I adore my Andromeda, but I was quiet reliant on StarXterminator when creating a minimized hydrogen mask for it. Just when I thought I had my processing down to mostly waiting on the CPU/GPU to crunch, you bring me back in!

Súper nice challenge

Download the data: github.com/deepskydetail/no.ai.challenge Submit here: deepskydetail.challenges@gmail dot com Discord: discord.gg/y2hgupCqfV Buymeacoffee link for .stl file: buymeacoffee.com/deepskydetail/e/355572 Step 9 should say "Added screen layer and adjusted levels to make background darker." (around 5:10)

please show the same parts of the images for comparison, what you did looks terrible, how can you compare when you move the image around the screen and don't show the same parts of the frames next to each other?

Just noticed that you have already posted a 300 vs 600 comparison. Thanks for this informative video. I am shooting 600 seconds over 300 seconds, but based on your statistical data, it doesn't seems like that i am gaining much. Thank you once again!

Thanks for sharing 🎉

Interesting. I try to adapt my sub-exposure length so I don't get a too bright night sky from surrounding light pollution sources, or to blow out details from brighter objects, which leave me often at or around the 15s mark. I'll try to max out my exposure length in the future.

Nice work! Glad to discover this channel. I was always wondering the same questions and I now understand better.

Love the rationale and reasoning before proceeding into the experiment. Your channel is a delight.

Wow, what an effort you put into this video. Subscribed 👍 You were mainly talking about the faintest areas in a stack. My main takeaways: .just overcome readnoise; doesn't matter if you take 300s or 600s subs. Stack in 32 Bit . What do you think about burned out areas ? For example M31. What would you do with Andromeda ? If you take longer exposures the burned out area increases. Or is the only way of working with those bright targets taking short & long exposures and try to combine them ?