Denoise Model V.2

Cannot find / select Denise Model V.2.
On the other hand, Sharpening Model V.2 is available (and impressive).
Why is that?

Are you talking about Remove Noise - Raw or non-Raw?

We released two new Remove Noise - Raw models, and a Sharpen - Standard v2 model.

RAW (Fujifilm RAF, lossless compressed, camera model X-S20, which is supported according to publicly available information). Computer: Macbook with M1 Apple Silicon. Stand alone use of Topaz.
As explained, for sharpening the new model (Standard v2) is selectable (and nice, at first glance), but for remove noise, no new model is selectable. Respectively, if you click on the “I” in a circle, it is stated that Raw Models “Raw Normal” and “Raw Strong” are available, but the app never gave the option to select “strong”. Maybe my images do not have enough noise (:smile:). But this is not the point. According to publicly available information, I have expected that a “v2” would also be selectable in “Remove Noise”. Misunderstanding?

Ah I see. If you are using a Fujifilm X-Trans camera then the Remove Noise - Raw models are not available.

We train our Raw models with Bayer sensor data. We do not have a special model for the X-Trans sensor unfortunately so it uses our Remove Noise - Linear Raw model which is not as good for noise reduction.

I would recommend converting the Raw to a TIF and then processing the TIF for better noise reduction.

I am adding this information to our documentation for other users to be aware of.

Dear Lingyu

Thanks for the reply.

Too bad to hear that about the X-Trans sensor.

That is the first bad news I hear about or from Topaz Photo AI. And this is not a minor point, quite the opposite.
Just days before v2 came out, I invested in a better Fujifilm camera body, and then this. If I only had known this before, then I might have decided differently, maybe switched to a Bayer sensor type camera.

Appreciate that Topaz will publicly document this now, but too late for me.

There is one thing I do not understand. The „old“ Model works quite well with the X-Trans sensor images, at least considerably better than the present No Noise AI of On1 Photo Raw 2023, or C1 Express for Fujifilm, or other apps I tried ( I have never compared with Lightroom, or DXO, but might be forced to do that in the future).

But OK, you said the „old“ model is „linear“. And the new model apparently is something else, apparently something non linear. I do not pretend that I have sufficient technical knowledge to understand this, however.

No hope that this sensor will ever be supported by the new Remove Noise Model, respectively a variation of the same?
Is the market share of Fujifilm just too minor for making this effort?

Kind regards,

Thomas Meili

The data in a Bayer color filter array looks like this:

The data in an X-Trans sensor is like this:

It would require us to train an entirely new machine learning model specifically for this different sensor as the data is different. This training is research intensive and may require a whole new approach. We haven’t dedicated time yet to explore the option.

It would be great if we could support it in the near future. At the moment, there are other tasks which are higher priority.

I hope that helps.

Dear Lingyu

I have noted that he following is stated now under the heading „supported cameras"

Fujifilm X-Trans cameras cannot use Remove Noise - Raw models due to the non-Bayer sensor data.”"

However, I believe this statement raises as many questions as it answers, as follows:

  • I bought Topaz AI beginning of this year because I was impressed by the remove noise feature - with images of the x-trans sensor. This applies to the now „legacy“ remove noise function, which is still available in v2, and still seems to work the same way which I still like.

  • The „legacy“ remove noise function - with options like „strong“ - uses „Remove Noise - Raw Models“ as well, doesn’t it? As you have stated, those „legacy“ models are „linear“, as opposed to the somewhat more AI driven “non linear” models of v2. Therefore, the statement is over-inclusive in that it does not differentiate between the 2 different sets of models, isn’t it?

  • However, this seemingly over-inclusive statement raises one concern: Does this mean that the „legacy“ models will ultimately - rather sooner than later - go away, meaning my investment - as a user of a Fujifilm camera - in Topaz AI will be gone for good?

Therefore, please clarify at least for me and in particular with regard to the last question.
And a last question: Am I the only Fujifilm camera user who works with Topaz AI and who may be concerned?

Thomas Meili

Currently Fujifilm RAF files from X-Trans cameras will use the Remove Noise - Linear model. This is the one you have been using so far.

Normal Raw files use the Remove Noise - Raw Normal v1/v2 and Raw Strong v1/v2 models.

I am checking with my team if the noise reduction would actually be better with the Remove Noise - non-Raw models (Normal, Strong, Extreme). If we determine that the non-Raw models do a better job than the Remove Noise - Linear model then we will switch to the non-Raw models.

However, this is only if the result is better with the non-Raw models otherwise there is no point in making the switch. So in the end we are still trying to provide the best noise reduction result and you should not see decrease in the output quality.

Dear Lingyu

Thank you for the quick reply. However, your reply seems to miss my points:

  1. The general statement on the Topaz website that Remove Noise Models cannot be applied to X-Trans files goes too far. The statement should specify which models exactly cannot be applied. My understanding was that this would be the new v2 models, not the v1 models, as the latter are still available in Topaz AI and work well with X-Trans raw files.

Again, I was and still am until now quite satisfied with results achievable with the v1 models (even though DXO DeepPrime DX which explicitly supports X-Trans-files, provides somewhat better results at the margins).

  1. Due to the generality of that statement, and now also due to the fact that your present reply ignores this issue (you only say what is „currently“ the fact, which I am perfectly aware of), I am concerned that the v1-„linear models“ for Raw might not be available at some point in the future anymore in the version(s) vor which I have bought a perpetual license.

Therefore, please specify whether or not these v1 models will continue to be available in the future or not.

Thomas Meili

RAF files use the Remove Noise - Linear model which has not been updated for a long time. Currently they are locked to this one model and do not have access to any of the Remove Noise - Raw models.

Raw files that are Bayer CFA have the following models for Remove Noise.

I am currently testing if the Remove Noise - non-Raw models would be better than the Linear model. If it is, then yes I would want to remove the Linear model as the other models are performing better denoising.

If the Linear model was still comparable to the non-Raw models, then we would keep it in future versions. We will not remove the Linear model from past versions which you already own as we do not go back to modify previous versions.

Thank you for your reply.

My understanding of your explanation, the RAW Remove Noise Linear Models might go away in future Topaz Photo AI versions, but only if you find that the Non Raw v2 Models are better, and if this option is provided for RAW Files as well.

Therefore, you have confirmed that the statement on the Topaz Webpage, according to which none of the Remove Raw Models can be applied to X-Trans-Files, is wrong.
I just recommend to clarify this statement in order to avoid gross misunderstandings by the Fujfilm user. That would also benefit Topaz, would it not? Unless, of course, you want to get rid of Fujfilm X camera users.


Thomas Meili

Here’s the situation as clearly as I can state it:

Linear Raw and Fujifilm RAF files both use the Remove Noise - Linear model. They cannot use the Remove Noise - Raw models.

I am testing Linear Raw and Fujifilm RAF files with the Remove Noise - non-Raw models to see if the results are better. So far, it does look like the Remove Noise - non-Raw models will have better results in almost every case.

Since the results are better, I am going to try removing the Remove Noise - Linear model and use the non-Raw models for Linear Raw files and Fujifilm RAF files.

The Linear model is not using Raw sensor data, which differentiates it from the Raw models.

Dear Lingyu

Many thanks for your relentless efforts.

Apparently, based on this information, my understanding of how denoising by Topaz - and maybe competitors - works was fundamentally wrong.

According to your information, the remove noise function of Topaz Photo AI, and the Topaz DeNoise App, has indeed n e v e r - also prior to introduction of v2 - used the “raw sensor data” if you applied it to a RAF file.


This is now also clearly stated on the features website and in the app itself. If you click on the (i) next to the „AI Model“ under „Removing Noise", it is stated now that the Raw Normal and Raw Strong models cannot be applied to this file.

My understanding therefore: The „single specialised model“ mentioned now in this (i) in the app is presently identical with the „linear model“ you have mentioned - but may possibly soon be replaced by the new non-raw models.

Thanks again, and regards

Thomas Meili

1 Like

Yes we are definitely on the same page. Thank you for having this conversation with me, I started testing the non-Raw models against the Linear models and added more information in the Topaz Photo AI articles based on this.

Please let me know if I can help with anything else. For now, I’ve extended your license by 3 months.

Thanks again.

Now unfortunately, I have to bother you with one last question: I purchased a perpetual license. What exactly does the extension by 3 months mean in this context?


Thomas Meili

The 3 month license extension is for upgrades. When you purchase Topaz Photo AI you own all versions released before that day and all updates released while your license is active.

Once your license is inactive, you still own those previous versions but receive no new versions.

OK. So, thanks again.

Thomas Meili

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.