The effect of resolution on StarNet++

[SkyHiker]

For those using Starnet++, I found that the results can be dramatically different depending on what resolution is used. I had a NB image of the center of the Heart nebula at 6249x4176 that took about an hour to process with StarNet++.

The result was not good - splotchy. Click on the image to see how splotchy it is.

BTW the original image had plenty of noise, especially since I only had 2x5 min SII subs (the other ones were 10x5 min) so I had to crank the heck out of it and reduce the blown up stars before I got the above image. The processed SII image had a ton of noise, which you can see.

Later I rescaled the image to speed things up to 1000x668 and got a much better result albeit at noticeably reduced resolution (but not a whole lot, and it is so much quicker).

Then I went on to 2000x1337, this is the one I like best:

and 3000x2005, here the large-star imprefections begin to show:

Bottom line, the 1000 image processes the large stars best but suffers from reduced resolution. The 2000 image fixes the resolution problem and shows minor artifacts from the large stars. The 3000 image increases the resolution but starts to suffer even more from large star artifacts. The original-size 6249 result is garbage (and takes forever to process).

I also played with the stride parameter. As it turns out, 64 is a good value. But if you are in a hurry, try 128 because the difference could be unnoticeable but it is 4x faster. If you get about at least 400 tiles you should be OK, 1000 is better.

Hopefully this is useful for those complaining about StarNet++. Apparently you can get much better results much faster if you reduce the resolution! StarNet++ is free and can be used without depending on a plugin program such as PS, PI or Gimp.

[yobbo89]

very nice, i have a beefy computer, i can't remeber the time but i think it was like 20 mins for a 2 pannel image or something ... alot quicker then your stock results but still slow, 2 coffees later . have you tried drizzling after ?

[SkyHiker]

very nice, i have a beefy computer, i can't remeber the time but i think it was like 20 mins for a 2 pannel image or something ... alot quicker then your stock results but still slow, 2 coffees later . have you tried drizzling after ?

I have a Toshiba L775 laptop from 2011, a Core i3-2310M with 2 cores/4 threads and 6 GB of RAM, maybe that's why it it takes longer. No drizzling here, my system automatically prevents that (I tried it once and it ran out of memory).

I bet the resolution aspect has something to do with the size of the images that were used to train the neural net. I'll dig up some other images and see if I see the same effect.

[STEVE333]

Nice comparison Henk -

Even though Starnet leaves artifacts of big stars those artifacts often disappear when the stars are added back in.

If the artifacts remain a problem even after adding the stars back in I've sometimes had success by using Clone Stamping to "heal" the artifacts in the starless image before adding the stars back in.

Just FYI