11.07
#...
·
3
likes
|
|---|
|
I noticed on the recent Topaz update that a new "Severe Noise" model had been added to the AI product. I had all but stopped using Topaz AI due to star butchering, invented details, and odd shadows patterns so I was interested. I opened an RGB image and pumped up the vibrance and saturation to get some really nice noise going. For once when I turned of sharpening in Topaz it seems to have not sharpened! Things look pretty good here. I found that applying even a little sharpening and the small stars were butchered again, but I'm really not looking for Topaz to sharpen anyway. Let me know if you are seeing worthwhile results with your data using this new model.  |
11.57
#...
·
3
likes
|
|---|
| Looks interesting. In the previous versions of Topaz DeNoise, I always think it "creates" too much details that don't actually exist and sharpens the image too much. If the new version solves this, I will consider to use it. |
2.71
#... |
|---|
| Yes, it's promising! |
10.79
#... |
|---|
| @Kevin Morefield could you post a side-by-side comparison of the same image area? Certainly one can see the effect on the noise but I think one could give a better judgement if one can also observe the effects on the very same data. Personally, I'd be totally fine the with quality of the noisy image as the noise seems to be free of any systematic pattern, spikes and isn't too intense. |
8.31
#... |
|---|
|
Thanks, Kevin! I think the result looks very nice. No harm to the stars and the background is rendered in a very natural way. -Min |
|
11.07
#...
·
3
likes
|
|---|
@Kevin Morefield could you post a side-by-side comparison of the same image area? OK here are some crops of before and after. This is RGB data, unprocessed except for ABE, a stretch and the aggressive vibrance and saturation to enhance the noise.   |
|
10.79
#... |
|---|
|
Thanks, Kevin for posting the two images! Here are my observations: The high frequency color and luminance noise is reduced. In the Topaz image, the patchiness of the background becomes more prominent. The spiral arms and fainter stars and objects seem to loose color (e.g., the outer rims of the spiral arms appear to have some slight purple cast which is gone. The faint objects around the bright star top left to the galaxy center become gray). |
|
2.71
#... |
|---|
|
Here's some very fast testing. trying to upload as PNGs, so we'll see. 1) Original, fairly noisy and stretched stack. (something with a very grainy background for testing) 2) "clear" mode in auto (high noise removal, low sharpness) 3) "low light" mode in auto (70,39) 4) "low light" mode, 70 denoise, 0 sharpening 5) "severe" mode in auto (70,39) 6) "severe" mode, 70 denoise, 1 sharpening (lowest possible setting) HH.zip |
|
2.71
#... |
|---|
|
Original, low light (smooth only), severe (auto), severe (sharpen nearly off) HH.png HH-lowlight-(70,0).png HH-severe-auto(70,39).png HH-severe(70,1).png (Again, a very stretched, noisy example!) |
1.81
#... |
|---|
In the Topaz image, the patchiness of the background becomes more prominent. This is perhaps simply because the high frequency noise got removed? |
|
10.79
#... |
|---|
Yizhou Zhang: Yes, in deed. If the high frequency part is damped, you'll see the low frequency part more clearly. The patchiness is also there in the noisy image (just turn up the brightness of the screen and don't zoom in to see the whole image). The high frequency noise "distracts" from it. In the end, the balance matters. With some of the high frequency noise, one can have a more homogeneous background than without. |
|
11.07
#... |
|---|
Yizhou Zhang: I agree. I typically apply noise reduction like this with either a mask or opacity of about 80% Leaving 20% of the noise seems to maintain the integrity of the image. |
10.16
#... |
|---|
| I have never tried Topaz. But looking at the pictures above, I wonder wether the denoising is really better than the job Photoshop or Affinity Photo could do. How big is the difference? |
|
10.79
#...
·
1
like
|
|---|
Olaf Fritsche: if the Topaz model effectively behaves line a low pass filter then likely there won’t be a significant difference. The only and somehow major difference would be that one knows the behavior of a low pass filter, while the „AI“ model may produce unforeseen effects if it is applied to data which is inherently different than the training data upon which the Topaz model is built. |
6.80
#... |
|---|
| As someone who has used PI much longer (TGVdenoise in both linear and non-linear space), but now experimenting with Topaz denoise, my first attempts to apply Topaz denoise on an already PI TGVdenoised image produced either no improvement, or worse artifacts & more noise (depending on settings). I'd be interested to hear other PI user perspectives on any additive difference(s) that Topaz denoise brings to the table for those already using PI's TGVdenoise.... |
9.48
#... |
|---|
|
Kevin, thanks for posting this information. I downloaded the latest version of Topaz AI and gave it a whirl. "I typically apply noise reduction like this with either a mask or opacity of about 80% Leaving 20% of the noise seems to maintain the integrity of the image." That's exactly how I have used Topaz AI on a few of my images, except I use a 50/50 blend of Topaz and standard noise reduction to achieve a natural-looking result that is somewhat less noisy the the standard NR permits by itself. Generally, I'm impressed with what Topaz can do for problematic noise in the sky background and the shadows of the images. It's great for dark/dusty nebulae, maybe better than anything else out there? However, I don't want Topaz touching my stars or the midtones/higlights of the image, where the S:N ratio is already good. The ability to turn off the AI sharpening completely is huge to me. I tried the newest version, and I found that the Severe Noise model with Sharpening = 0 still chews up the stars, but in a slightly different way. In the Topaz GUI, the stars develop a slight pixelation effect that resembles under-sampling, and they are distorted around the edges after the image is saved. I'm not sure it's an improvement over the previous versions. The smaller stars are totally destroyed in the few images I tried processing. |
0.00
#... |
|---|
|
My first foray into using this tool. I had been using the AI version apparently and was noticing lots of unwanted AI detail being added where it shouldn't be in the first few images I processed. I then noticed the new feature mentioned above and used it on my Wizard Nebula image and think it did a pretty nice job overall. I actually added back in some noise using the "original detail" option in Topaz to leave a small amount of grain and to keep from eroding small details. The conclusion I came to was that overall I think this software can produce some pleasing results so long as you watch carefully at what it's doing. I also reprocessed my Bubble Nebula image using the same features, but left a little more grain in this one. Edit: I processed both these images in Topaz with no stars, then added the stars back in Pi. |
|
11.07
#... |
|---|
Kevin, thanks for posting this information. I downloaded the latest version of Topaz AI and gave it a whirl. After the positive results with my NGC5566 data I tried the new model on M14. There were no settings where the stars were not chewed to bits! So clearly Topaz is not a solution for all images. As you mention, it doesn’t make sense to apply noise reduction to areas that contain sufficient signal sample to begin with. So my normal approach is to is use a reversed luminance mask to apply noise reduction. I usually adjust the mask to fully block stars and highlights and allow 80% to apply to the background. |
|
9.48
#... |
|---|
| I've used a "grayscale" mask like you are describing as well. It usually does the trick, especially if you play with the Levels/Curves on the mask until it looks just right. I got away with a simple 50% opacity setting in the top layer on a couple of images as well. |
22.40
#... |
|---|
|
Kevin, I'm aware of Topaz but I've never used it. You results look very nice but I'm curious how it compares with MureDenoise? John |
|
11.07
#...
·
1
like
|
|---|
John Hayes: John, I haven't used Mure yet so I can't say (on the list for sure). But till now, we've seen a number of imagers using Topaz, some with destructive results. I tried it early on but after actually looking at what it was doing I stopped. So I am interested to see new models being created that might be able to work with our images. FWIW, the Astrobin judges have had lengthy discussions about Topaz AI and the need to look closely at images being judged for tell-tail signs of the algorithm creating false detail and sharpness. Kevin |
9.15
#...
·
2
likes
|
|---|
| There is no reason not to apply MURE denoise as the first step of denoising. Even just a very light touch helps the image. Depending on how faint the target is, additional denoising applied to the linear image in PI (TGV or MLT denoise) may also be appropriate. Immediately after stretching the image is when I usually apply Topaz Denoise as non-linear denoising. I always remove stars before feeding the images to Topaz Denoise. Also, I use Topaz Denoise as a layer in Photoshop and then control opacity as well as masks if artifacts arise or if the effect is too strong. The stars can be added back either immediately or later with a Linear Dodge layer in Photoshop. |
5.88
#...
·
3
likes
|
|---|
|
Topaz has improved over time. I've had it for about a year but it messed with the images so bad I rarely used it. After seeing this thread.. I ran the below test on an already completed image. Left is with Topaz (severe noise option, remove noise = 4, all other options are 0). In my experiance you have to play with the settings.. settings that work great on one image don't work well on others. Other than the impact on stars.. my biggest complaint is the blotches you get in the backgrounds sometimes. For sharpening I like deconvolution in PI and unsharp mask in PS. For Noise reduction.. this seems to do a good job.  |
