Poor results with A.I. Clear

Hello again,

I am opening a separate thread for this issue as I don’t want to lump together too many topics. I’m also having separate issues with Gigapixel which I should be posting about soon. But for now I just wanted to share some feedback on A.I. Clear, in particular with some astro photos that I would like to process.

The image in question was shot on a Sony a7R II, continuous low mode (i.e. star eater not in effect), LENR OFF, ISO 640, f/2.8, 32 seconds. It was encoded as a lossless uncompressed raw file. The image was opened in Adobe Camera Raw and then saved as a TIFF, with only a custom camera profile applied, and no further adjustments, all settings zeroed, certainly no sharpening or noise reduction applied (which I understand can wreck havok with getting optimal results from A.I. Clear). I imported this TIFF into Studio (V1.13.1) and applied A.I. Clear on it, and also ran it through Topaz Denoise 6. Please refer to the attached JPEG, labelled to indicate the processing settings for each crop. A variety of results from different combinations of settings are shown to prove my point.

I immediately noticed that regardless of the settings, I was getting a dark contour halo around the starburst, and it is the same story for any other high contrast edges (brighter stars), sometimes getting worse or better depending on choice of ‘High’ or ‘Low’ for Enhanced Sharpness setting. The best possible result from A.I. Clear was using the ‘low’ Reduce Noise setting, with Recover Detail set to 1.00 (maximum), and Enchance Sharpness at Low, but it is still significantly worse in that regard than Denoise. This is most likely due to the built-in sharpening optimization that A.I. Clear adds. This isn’t something the user has any control over.

In addition, there are regions where the colour noise reduction was randomly not applied by A.I. Clear, so the result is a blob of colour around some stars.

In Topaz Denoise, a good (not necessarily optimal) result was obtained with the Strength set to 0.05, Color noise reduction at 0.20 and Recover Detail at 0.33. Note that the result is very nearly identical to the A.I. Clear Reduce Noise Low, Recover Detail 1.00, Enhanced Sharpness Low result in terms of overall amount of noise reduction, except it is much more uniformly applied and has almost no artifacts. Overall, it is a significantly better result than anything I could get from A.I. Clear.

I wish to quote from the help page of A.I. Clear:

Allow Dr. Partha Acharjee, a research scientist with Topaz Labs, explains it best:

“Recent advancements in Artificial Intelligence has shown great potential in image noise reduction. We were able to develop a proprietary neural network, (a type of AI that learns by studying millions of examples) architecture and realistic noise model to perform well on real-world, high-iso photos. It not only achieves much higher quality result than existing products, but does so automatically without the need for any manual tweaking. Moreover, our custom neural network, trained on millions of photos, can run on any GPU on both MAC and PC, which ensures faster processing on everyday computers without the need to upload your images to a cloud server.”

At this point, I am failing to see how A.I. Clear is achieving a “much higher quality result than existing products”. If anyone has any helpful suggestions on how I can achieve that, I am most willing to learn. Thank you!

1 Like

I see some differences but I think you should try these settings for A.I. Clear

  • Remove Noise - Med
  • Enhance Sharpness - Low
  • Recover Details - 0.10

Also something you should be aware of when you shoot long exposures is to turn DRO off, have Long Exposure NR turned Off and High ISO NR set to Normal.

I would also look at using DeNoise for something like this starting with the DeNoise 5 presets using the RAW Strong Preset as the starting point and then adjust the DeBanding by turning on both Horizontal and Vertical and adjust the Banding Width.

Hello again AiDon, thank you for your prompt replies to my posts. I tried your recommended settings but it turned out as I expected - it is among one of the least pleasing results from A.I. Clear, compared to the giant crop array above.

I’m afraid none of the Med or High settings for Remove Noise is any good in this case. The noise reduction is just too heavy handed, especially with a very low amount of Recover Details. Nothing I do would remove the ugly halos that develop around high contrast edges.

I did mention quite a fair number of details regarding the way the image was captured. I was shooting in raw, and the exposure mode was manual. I have DRO off, but it wouldn’t affect the final result because of this setup. If you read my post again you will see that I did mention LENR was off. High ISO NR was off but again it does not affect raw files.

That gives an absolutely garish result. I’m very familiar with Denoise and I can say that none of the presets offer anything of quality as the starting point.

That would only be useful if there is banding to remove. This is not the problem that I am having. I do appreciate your super fast replies but with all due respect, it might be more helpful if you could read my posts more carefully and address the relevant issues. I would really appreciate that. Thank you!

Raise a Technical Support Request at the Need Help link above.

As I mentioned in the other thread, I had done so and attempted several times to reach out to Topaz technical support in this way but my messages have all gone ignored. Hence why I am now posting publicly in this discussion forum.

So I take it you had no response from Tech Support?

Hi @JoeFedric-TL could you have a look and see if there are any Tech Support Requests from this guy’s email address?

Have you tried opening the RAW file in Studio, although it does give a Linear like conversion (flat) it may be able to give a comparison between the TIF from PS/Lightroom using ACR and it.

That was one of the first things I tried.

The tonality is almost exactly the same as the result from ACR, because I did not apply any adjustments in ACR - all settings zeroed. However, the demosaicing interpolation is significantly worse in Topaz Studio. The effects of A.I. Clear on opening a raw file are similar, apart from the worse demosaicing. This is to be expected.

Yes.

Please look at support request no. 150356.

We’re a small team. It’s important to understand that we cannot answer all requests immediately, and when we get a large number of requests, some people will have to wait. I’ll look into your request. Thanks.

1 Like

Hi Joe,

Thanks for responding here.

I absolutely understand. But my request was opened on 18 August, and I received your response on 22 August requesting additional information and images, which were provided by 23 August. It has since been over 3 months. Then after publicly posting for help here, I receive this response to that help request 10 hours later:

We’ve released a number of updates to the products you’re referencing here.

To start your trial, can you please try the steps and guidance provided in the article below:

Reset Local Database

Please let me know if that worked for you!

Firstly, I don’t think it takes over three months to provide this generic response. In the meantime I had figured out the problem and managed to activate the trial, otherwise how was I able to provide the results from A.I. Clear here?

Also, if you had taken the time to read my post, you would notice that I am using the latest version of Studio.

The problems that I am raising are critical bugs which affect all users of Studio, even in the latest version. The GUI preview is grossly different from the actual renderings being done by the processing engine, evident the moment one exports the file out of Studio as a TIFF and then comparing with what the preview is showing. I’m actually a little surprised that more Studio users are not complaining about this, or even raised the issue. It is impossible to judge the exact effects of each of the wonderful specialist editing tools that you and your team have worked so hard to program.

The additional feedback for A.I. Clear here might be helpful for the developers to understand how professional photographers with a trained eye for detail find the noise reduction and enhancement quality of A.I. Clear and how it could be better. At the moment, the pernicious detail restoration is too aggressive at even the “Low” setting for Enhance Sharpness. It would be best to have a way to turn it off entirely. It is destroying all the wonderful fine detail that the A.I. noise reduction process is working to hard to reveal, thus being totally counteractive to your goals.

Last night I tried A.I. Clear on another regular noisy landscape image which is not stars, rather a sunset at a lake. The effect of too-strong halos remain. In certain ways, it is arguable that A.I. Clear gave a superior result to Topaz Denoise, but in other important ways it has failed to live up to the claims. If I were making small JPEGs to show on Instagram, Clear gives a superficially better result, but if the goal is to make high quality fine prints, it does not hold up as well to scrutiny with all the grungy digital-like artifacts.

I am trying to provide useful feedback and I do believe the community of users of Studio will find this information to be helpful at least. I sincerely do hope that it is taken by the Topaz team just as well. I have been a Topaz plug-ins user for some time now, after being enamoured by Bart van der Wolf’s excellent sharing on the Luminous-Landscape forums. They are without a doubt first-rate plug-ins for high quality work, and I do desire to support companies that have been instrumental to my work and business. I also do believe that I can provide helpful feedback to for them improve their products, as I care about them and want them to continually be the best that they can be. I have previously always held Topaz with high regard and I hope to do so in the future. Please restore that confidence by being more attentive to the details. Thank you for your time.

I think it’s rather obvious by your own statements, as well as the public reception of the software, that what you describe is the exception, rather than the rule. Expecting your experience to mirror everyone else’s exactly is not a great way to view troubleshooting. Instead, understanding what the problem actually is by evaluating the output is most important.

Noted.

Can you provide some samples of this?

Care to elaborate on what you mean here?

How so?

It is possible to see artifacts in some places when you start with a high-quality image. The models were trained using low to medium quality images, rather than exemplary photos. The result is that A.I. Clear works much better with images that actually need considerable improvement.

We love the feedback our users provide us. However, often we receive explanations about how things are wrong or bad, but not given much more information than that. We do our best to figure out what, exactly, the complaint may be about, and try to identify the cause to hopefully implement some sort of fix or feature that addresses that complaint.

That said, feedback is implemented on a development schedule, so it is reasonable to allow for time to pass before your particular issue is addressed, unless you’re raising a product-breaking or product-wide problem to us.

I do acknowledge that with complex systems, there are apt to be errors/shortcomings/problems which pass by almost everyone. Otherwise, rockets would always take off perfectly and software would never need bug fixes. The preview mismatch will likely be invisible to anyone who is looking at 100% zoom or less, which is probably the case why no one else has seemed to notice it. But at higher than 100% zoom, since Studio does not allow us to set the zoom percentage in whole 100% steps, and it does not use nearest neighbour interpolation (that would likely be worse at non-whole steps, even Photoshop doesn’t do that except for 200%, 300%, 400% etc.) it already causes significant degradation (blurring) of the preview.

The degree to which it is obvious or subtle is also relative of course. For professionals doing high-quality work and making large detailed prints, we tend to view what is inconsequential flaws as more pernicious.

Yes, that is what I have been trying to do with my posts here, and also originally trying to reach out to you by email. When you did not respond to me, you could not possibly understand what the problem is.

To provide further evidence to support my case, take a look at this new jpeg, which shows side-by-side the screenshot of the preview window and settings in Studio with A.I. Clear adjustment applied, and the same part of the image output as a TIFF (which had to be converted into JPEG to upload here). The left column is 100% magnification, and as I mentioned, it is very difficult to see any difference at all. But when enlarged 400%, one can see the cross-hatching pattern in the preview as well as the more prominent halos.

Note that Studio’s preview does not have more prominent halos when using the Basic adjustment, and the tonality and colour are an exact match to the TIFF output. This only happens with A.I. Clear. However, the cross-hatching patterning affects the preview all the time, and is especially obvious when zoom magnifications are over 100%. Looking at smooth regions of an images e.g. sky, makes it more noticeable.

Now, this effect is a little more pronounced when using the zoom tool in Studio itself, except that I am unable to set exactly 400% with the zoom tool. Here it is at 398.6%. It is a serious issue for anyone wanting to judge absolute pixel quality of their images.

Thanks for explaining that the goal of A.I. Clear was not intended for producing flawless, top-notch work, and indeed such work should not ever require such heavy-handed rescue with editing software. Nonetheless mistakes happen (like with the Hubble), and sometimes extreme methods are needed to rescue a precious picture which cannot be made again. I am happy to lower my standards in such a situation.

Yes. Compare the original to the A.I. Clear version, there are now visible halos on the reeds. Processing settings are as labelled. No combination of settings for A.I. Clear thoroughly removes the halos. They are just more or less visible. This example shows what the ‘auto’ result is.

Now compare A.I. Clear with Topaz Denoise 6. No amount of adjustment with Denoise will ever help me achieve the beautiful clean look that A.I. Clear is giving (low noise plus sharp details). Math-based processing is unable to replace detail in the original image the way A.I. processing methods can, simple as that. So there is no accounting for taste, but I believe the majority of people will prefer the sharper, crisper, cleaner result from A.I. Clear as a web jpeg such as this. However, the sharpening halos and the somewhat crunchy-looking reeds give the impression of digital artifacts and could be better. There are also color halos around the reeds (just as there was around the stars). Turning up the ‘Recover Details’ slider all the way to 1.00 goes a long way to making the halos almost not visible, but the color fringing halos still remain around the higher contrast edges. It is arguable that Denoise gives a more organic or analogue effect. Also the result can be further sharpened without fear of exaggerating any halos. It is difficult to pick a clear winner (pun not intended). In this case I would probably combine with hand-painted masks both renders in Photoshop, to get the best of both worlds. But it would be nice to just have an “OFF” switch for sharpening in A.I. Clear!

It goes both ways. I went to lengths to provide details and examples as best as I can in my email, but when they were ignored for 3 months, it leaves me the customer wondering if the company cares. As it stands, it is evident to me that if I did not raise these issues publicly here, I would not have gotten a response.

I would like to make it plain that I never once demanded anything be fixed right away, only that my feedback is acknowledged. A simple, gemeric message like “I’ve received your email. We are currently very busy attending to other customers requests that have come in before yours, but we will get back to you in XX working days”, goes a long way to reassuring your customers that you are still looking into it, especially after your first reply to request for further information. Don’t leave us hanging! It’s just nice customer service. As Head of Support, I hope you can share this with your team. I am sure all of us Topaz customers will be highly appreciative of such acknowledgement. :slight_smile:

Have a great day everyone,
Sam

To add on to my previous post: I brightened the sunset picture before running AI Clear and Denoise on it so the difference would be more visible on screen. But here it is in its unbrightened state (which is the current choice for my editing, so I can keep most of the ugly noise dark and nearly invisible in the print). Note now that AI Clear is struggling to re-create detail in the reeds, and a significant amount of ringing artifacts are present, as well as the halos like before. AI Clear is now definitely worse than Denoise. The best final result might be a blend of the two (using the smooth regions from AI Clear), but AI Clear has also reduced the saturation of the colour of the water too much.

Here is another example towards my case. Here is the full view of the image in question, which is slightly diffused sunlight reflecting off some debris on a pond’s surface. That region is of exceptionally high local contrast, while the foliage backdrop is of exceptionally low contrast. It is difficult to balance the degree of noise reduction without too much loss of detail. A worthy challenge for Denoise and AI Clear.

Here are the crops of the original, best result I could achieve with Denoise, and AI Clear. AI Clear is noticeably worse off in this example. There are many colour botches which were added around the light sparkles which were not there before, and the halos again are destroying fine detail. Increasing the Recover Detail slider from the default of 0.1 to >0.5 helps to recover only marginally better detail in the low contrast foliage, but the ugly colour blobs remain just as visible. As with the previous examples. AI Clear fails to produce a comparable result to Denoise on fine detail, and generates ugly artifacts around higher contrast edges consistently.

Edit: I urge anyone looking at these samples to not just open them in your browser window, but to right click>save image as and download them to your machine and view them at 100% (or 200%/300% especially if you are on ‘retina’ monitors) magnification to be able to judge these issues fairly.

I have consolidated my observations made here over on the Luminous-Landscape forums. In there, some of the descriptions of the issues I am facing are more accurately described than I have done here. Updating this thread with a link for completeness.