Thanks a lot Karsten!
Unfortunately, there’s no studio version available for my system. Anyway, tried a clean install, to no avail. I’ll try to go back to the previous driver version tomorrow. Thanks!
My LR 6.2 doesn’t handle CR3 files so I first covert to TIF.
I did run Denoise stand alone and had it process the same CR3 files, then render as TIF. The results were very different from the Canon Digital Photo Professional conversion to TIF as the Denoise was brighter and “flat” in comparison.
My mistake, try the last whql.
Thanks for the observation. I’m still noticing differences between CPU and GPU. Not huge, but some variation in noise granularity and the shift in sharpness acuity are there. In all honesty, though, the differences between CPU and GPU are by far less discernible in Denoise AI compared to Sharpen AI.
Just a quick comparison between DXO Deep Prime (NR) and Topaz Denoise AI. I used the result of a test from Thomas Fitzgerald website where he tested Deep Prime on a very noisy night photo. I took his original image crop and ran it through Denoise. While Deep Prime did a credible job it still mushes a lot of detail that Denoise recovered (low light mode).
2 Likes
I spent some time with the latest Denoise and Sharpen. I ran both as stand alone and processed a Canon CR3 RAW file. I selected a processing method in each program (using the “auto” setting) then exported as TIF.
In each program I used the GPU and CPU processors, then pixel peeked at 200% and 300%. In my opinion, the GPU processed file was slightly softer than that from the CPU. But the CPU image had a bit of a “crunchy” quality. I’ll try rerunning the CPU method but manually lower the sharpening a bit to see if the crunch diminishes.
This was true for both Denoise and Sharpen.
Semantics aside, I believe this is exactly what I’m seeing as far as the difference between CPU and GPU is concerned. Still, the output after CPU is for me superior, but it takes hell lot longer to get it.
1 Like
Deep Prime already has an advantage.
What is particularly noticeable are of course the colors, where Denoise has problems, and details like fonts.
I also remembered it worse (Deep Prime).
Just because the rest of the software tends to get in the way of my image editing.
I think you guys are doing flawed comparisons. This is like testing colors rendition in RAW files. The user input and ability to adjust is the critical component and this is where most of these types of tests fail, because they don’t rely much on ability to adjust the image by skilled user.
Also every image is going to be context sensitive. For example, Topaz has serious problems with lot of color noise when camera sensor is stretched to the limit. This is where DeepPrime does amazing job and topaz does not. to some Extent this can be mitigated by first applying color noise removal in programs such as Lightroom or Capture one and than applying only luminance noise reduction in Topaz, but at the expense of not being able to apply the de noising in RAW.
Speaking of RAW de noising, its another problem since Topaz is not really designed for it. There is no correction of lens distortions and color rendition is something one cannot adjust and when you export DNG file its not fully readable as original RAW by other programs. So this means you have to use Topaz de noise at the end on a TIFF or JPEG, adding extra step, and limiting some of the features like shadow recovery.
Secondly, build into the De-noise from topaz are sliders for sharpening etc. However, the best workflow I found is DXO for everything regarding RAW conversion, including de noising, and than applying sharpening to TIFF or JPEG using Sharpen AI from Topaz.
In other words its the combination or tools and workflow that produces best results. Doing a random set and forget comparison like this on a random file will produce all kinds of results and one can choose to use which ever one thinks they like better as a software. This leads to confirmation bias, quite present among few people on this forum. Where Topaz can do no wrong.
Another benefit of Topaz is that works with JPEG or TIFF’s and not only supported RAW files like DXO. Which means that if you are Fuji shooter or working on unsupported RAW file sor just scanned files etc. Topaz is only option, DXO is not even available. But if you have supported RAW files DXO deep prime + Topaz Sharpen, not de noise produces by far the best overall results. Point being that they are just tools. Depending on the skill of the user, particular image etc. Results will vary and its not about which tool but results on the end.
Where one can compare the two tools is features and workflow advantages and limitations.
4 Likes
“What is particularly noticeable are of course the colors, where Denoise has problems, and details like fonts.”
Yes, I think most Topaz models try to invent details while some other AI programs like DXO tries to not invent just clean up. The workaround would be to use two models from de noise and than layer the results and mask it out in Photoshop.
For example using Low Light mode for everything not with fine details like text in the image. And than AI clear mode for the parts the Low Light destroyed. By combining the two, its possible to get quite good results even in challenging situations. The biggest problem I think with Topaz De noise algorithms is the high degree of color noise when you push camera sensors to the max. Than Topaz leaves ugly blotches of color everywhere and moving color noise slider only destaurates colors in the image, indiscriminately. This is something DXO does far better job off. Even Photoshop workflow with Topaz is not as good. Hopefully Topaz will fix that in the upcoming updates. its been around for last two versions.
P.S.
Topaz Sharpen and DXO allow adding back some texture by introducing Film Grain. Useful at times, and it would be welcome addition to Topaz Denoise as well to cut down on the number of applications one must use.
3 Likes
I ran my quick comparison only because CR3 readability was recently added. I agree with itunes663 “In other words its the combination or tools and workflow that produces best results.” From my Canon 90D every CR3 file must be edited. But it helps to start with a good conversion to TIF. For example, the latest Digital Photo Professional made a big improvement over the version only two sub versions ago.
+1, Exactly 
1 Like
I fully agree!
2 Likes
I’m not at all happy with the latest Denoise AI engine - for me, V2 worked a lot better.
How can I downgrade to the latest V2 version?
You can find 2.4.2 here:
and 2.3.6 here:
You could also try 3.0.2. I find the LowLight-Model in 3.0.2 very good in comparison to 3.0.3.
1 Like
Many thanks, Fraenzken!
I don’t think going back to 3.0.2 will solve my current dissatisfaction. As I mentioned before, for my application (bird images) Clear AI provides better results then Denoise AI. Plus, in V2 I was able to create a composite image consisting of a Denoise/Sharpen-ed layer with a crisp subject, and a layer using a 70-100 Remove Noise setting to blur out the background. Such as this example. This does not work in V3 anymore, because highler levels of the Remove Noise slider have a lot less effect.
Overall, I don’t think V3 was a step in the right direction if it comes to IQ. DxO Deep Prime provides really great results as a Denoise competitor. If Topaz will not be able to catch-up I will not probably not buy another year of updates.
3 Likes
Hi adam.mains,
I likewise tested and compared Deep Prime and DeNoise 3.0.3 thoroughly and concluded that Deep Prime does provide sharper and more convincing detail, perhaps Deep Prime does this by taking valuable parameters from the camera data and working with the AI model, but it also has the obvious disadvantage that if it is not the camera’s RAW photos, it won’t work. This is where DeNoise has the advantage that it works even when processing a low quality web image. My suggestion for future versions would be to add a more sophisticated standalone AI model to determine the state of the image, such as compression rate / from camera or phone / people, architecture, plants, paintings or landscapes / presence of faces, limbs, hair, etc. This predetermination would provide more precise and targeted adjustments, and could also be used for other Topaz AI products
2 Likes
Apart from the fact that DxO Photo 4 tells me how to process an image, because everything except the local corrections is done using sliders (and sliders are nothing more than mathematical defaults), the results are comparably good.
I used Clear (high), which suited the image best.
EOS R5, ISO 1250.
Denoise AI 3.0.3 (AI Clear) → Photoshop
DxO Deep Prime & its Processing
AI Clear (High) & Photoshop (Quick edit)
DxO
AI Clear & Photoshop
Original
2 Likes
