I have been using Topaz Photo quite actively for a long time, often comparing different models and combining them with Photoshop. Over time, I have accumulated several ideas that, in my opinion, could make the program much more flexible and predictable. Some concern the quality of the processing itself, while others relate to the interface and workflow.
I will start with the ideas I consider most important and the ones that would have the greatest impact on my everyday workflow.
- Frequency Separation
For a program of this kind, I consider this one of the most necessary features. A standard Opacity slider is often not enough: it reduces both the useful and the unwanted parts of an effect equally. In Photoshop, I constantly work around this with frequency separation. I subtract the processed result using Apply Image > Subtract, then use High Pass and Gaussian Blur to limit the effect to the frequency range I actually need.
This would be useful for almost any enhancement. Recover Faces 3, for example, could be applied more strongly to plastic-looking skin from a poor camera, then restricted so that it does not affect larger facial details. The model could restore skin texture without making the eyebrows, eyes, and other features look overly AI-generated. Super Focus v3 becomes an excellent low-frequency denoiser when its effect is frequency-limited: it removes blotches, dirty tonal transitions, and large-scale unevenness without adding unnecessary sharpness.
I consider this especially important because many users do not work in Photoshop, and even fewer know that this kind of control is possible at all. In Topaz Photo, it could be a simple universal control for any enhancement: instead of reducing the entire effect, the user could choose which spatial frequencies the model is allowed to affect. The control should also scale automatically with image resolution. The same setting should not behave differently on a 12-megapixel image and a 100-megapixel image.
One possible implementation would be two additional controls, something like Fine Detail Protection and Large Structure Protection. Internally, the first could control the radius of a Gaussian Blur applied only to the model’s change map, not to the image itself. The higher the value, the less the processing affects fine texture and sharp detail. The second could use High Pass to limit the model’s influence on large forms, tonal transitions, and the overall image structure.
There is no need to expose specific Gaussian Blur or High Pass radii to the user. Two simple and predictable controls would be enough to confine the model’s effect to the desired frequency range.
- Luminosity Only and Color Only modes
For enhancements where it makes sense, there should be an option to limit processing to luminance or color only.
Sharpen, Denoise, and other tools often change not just luminance but color as well. This can create colored digital artifacts along high-contrast edges, which then have to be removed manually. These enhancements need a Luminosity Only mode so that they affect brightness without altering color.
For Balance Color and other color-correction tools, the opposite is useful: a Color Only mode. Sometimes I want to correct the color balance without changing luminance.
A similar idea is already used to some extent in Adjust Lighting by disabling Color Correction, although for some reason the color still changes slightly.
- Lab color space support
I work in Lab fairly often, and Topaz Photo currently does not support it through the Photoshop Automation plugin. Ideally, I would like full Lab support through the Automation plugin. Until then, there should at least be a clear message telling the user to convert the image to RGB.
At the moment, if I launch the plugin from Lab out of habit, the plugin stops working even after I later convert the image to RGB, and I have to restart Photoshop. So the lack of Lab support also turns into a rather unpleasant bug.
- Redefine
A generative model similar to Wonder, but without upscaling.
This kind of model is needed for images that already have sufficient resolution: iPhone photos, digicam images, files damaged by overly aggressive smoothing, and images with heavy oversharpening. It should do more than enhance detail. It should reconstruct natural texture while removing characteristic digital damage such as ringing, oversharpening halos, and plastic-looking texture caused by excessive denoising.
Redefine Creative already exists, but it redraws details too freely. What I am suggesting is a fundamentally different model that should neither invent content nor upscale the image. The input and output dimensions should remain identical. The model should not alter geometry, face shape, object boundaries, color, or overall tonality. Its job would be to remove digital dirt and mush while restoring natural texture strictly within the existing image structure. In other words, it could reconstruct texture inside existing forms, but it should remain tightly constrained by the source and should not create new details.
Honestly, I am not entirely sure what the Recover models are supposed to do if not something close to this. In theory they should address a similar problem, but in practice they are very slow and rarely produce a truly convincing result.
My current workaround is rather strange: first I add strong Grain in Camera Raw until it hides the oversharpening and plastic texture, then I apply Denoise Max. I would much rather have a dedicated model that does this properly in a single step.
- Speed and proper masks
Image analysis should run in the background while the user gets immediate access to the enhancements, as in Gigapixel. I do not always need noise analysis or automatic subject detection. Sometimes I simply want to launch Super Focus v3 and start working, rather than spend several seconds watching the program make decisions for me.
The masking system also needs major improvement. Camera Raw often detects objects more accurately, and Topaz Photo’s masks are also noticeably slower, especially Superpixel. It would be better to let users disable live mask rendering and update the mask only after the action is complete.
There should also be a simple Never Auto-Select option in the settings, so an enhancement applies to the whole image by default unless the user explicitly chooses a mask.
I would also very much like to see the best ideas from Topaz ReMask return: exporting a mask to Photoshop as a proper layer mask, and the same boundary-refinement approach where the user roughly marks what must be kept, what must be removed, and the program recalculates the selection. The old ReMask did this quickly and surprisingly well.
- Rethink denoising
RAW Denoise itself is not bad, but the current set of sliders somewhat contradicts the basic idea of Topaz. The user should choose the result they want, rather than worry that one careless slider movement will turn the image into mush. The controls should be safe and predictable.
I would use four buttons: Off, Color Only, Clean, and Full.
Color Only would remove chroma noise, moiré, and chromatic aberration without touching the luminance component at all.
Clean would not simply remove noise, but refine it: eliminate coarse digital grit, unevenness, and blotches while preserving the character of the texture, so the image does not look “cleaned and then re-grained.”
Full would aim for the most complete possible noise removal without producing a plastic look. A note here: there is no fundamental limitation that makes it impossible to remove almost all noise from a RAW file without turning the image into mush. That is a question of model quality and approach, not an unavoidable trade-off.
I am not convinced that a continuous slider between Clean and Full is necessary. When I process a single image manually, I can already create two versions and blend them later using frequency separation. When I need to process a thousand wedding photos, I want to choose the expected result blindly and move on, rather than inspect every frame hoping the denoiser has not eaten hair, grass, fabric, or some other fine texture.
There is another separate issue: after RAW Denoise, the image geometry changes slightly even when lens corrections are disabled. This makes proper pixel-level comparison impossible.
I would also seriously rethink the regular Denoise enhancement. Even the old Denoise AI 2021 performs better in some cases. There are currently three models, all of them fairly mediocre, and it is almost impossible to know in advance which one to choose for a particular image. I would rather have one truly good and predictable model for photographs, plus a separate model for artwork and scans, which effectively does not exist at the moment. Strength could then be controlled with one overall parameter and the frequency control suggested above.
The following suggestions are less important. They would not change my core workflow as much, but they could still improve the program significantly and solve several specific problems.
- Update Balance Color
At the moment, I am more likely to use Auto in Camera Raw than Balance Color. Camera Raw is faster and often more predictable. In my opinion, Balance Color can be slow if necessary, but it should be a genuinely accurate automatic white-balance tool.
A promising direction would be a scene-guided model that takes lighting into account. It could recognize skies, clouds, snow, skin, foliage, and neutral surfaces, then use them as probabilistic color references while accounting for the illumination on each one. For example, the model should not try to force neutral objects in shadow toward warm gray.
I would remove automatic object selection from Balance Color. Here it tends to get in the way more than it helps.
The Temperature slider also seems like a poor fit for this kind of tool. Instead of asking the user to manually shift the temperature, it would be more useful to provide a simple Cooler – Neutral – Warmer control. The model would first find the most natural balance, then the user would only indicate the desired tendency. Topaz Photo should reduce the need for users to guess the internal parameters of the model.
- Completely rethink Sharpen and build something like EQ Sharp
a) There are too many models now, while only a few are genuinely useful in my work. Standard is the most stable and behaves roughly like an intelligent unsharp mask. Refocus can sometimes reconstruct detail well. Wildlife is also decent. Most of the others create too much mess too often. Even some of the official demonstration images look overprocessed.
I still do not really understand Portrait. In some cases, the effect seems almost opposite to what I would expect: if I subtract the result from the original and invert the difference, the perceived detail can actually increase, even though the model itself is slow.
One of the best sharpening workflows I currently get is through Gigapixel at 1x: Standard > High Fidelity > Frequency Separation in Photoshop.
b) The idea behind EQ Sharp is to work by spatial-frequency bands. I would use three models.
The first would enhance high-frequency sharpness.
The second would reduce excessive contrast in the upper mid frequencies when needed, creating something similar to a controlled glamour-glow effect.
The third would increase local contrast (clarity) without creating the harsh, over-defined look that excessive clarity often produces.
For high-frequency sharpening, there should be one or two model modes, three at most, to avoid unnecessary variation.
Photos posted to social media are often reduced to around one megapixel, and much of the fine detail is destroyed by compression anyway. Yet even after that reduction, a photo from a good camera usually still looks sharper than an iPhone image. The purpose of EQ Sharp would be to reduce exactly that gap - distribute contrast correctly across spatial frequencies and preserve the perception of sharpness where ordinary sharpening becomes almost invisible after compression.
- DeHDR / Tone Restore
Adjust Lighting is fairly slow, and I do not find the result very impressive. It is not suitable for blind batch processing. A truly automatic image-processing model would be extremely difficult to build and, in my opinion, probably does not align very well with Topaz’s main strengths.
A better idea might be DeHDR. It would remove the effects of aggressive computational photography and poor tone mapping: compressed highlights and shadows, hard halos, and flat, asphalt-like faces.
DeHDR could have two outputs.
Neutral would return the most natural possible raw-like, slightly flat base for re-editing.
Natural would first remove the previous aggressive tone mapping, then automatically restore a normal distribution of light. The useful parts of Adjust Lighting could be integrated here.
- Clean Max: A Model for Scans
A reference to Topaz Clean 3.
I would like a dedicated model that turns a scanned or photographed page into a clean digital version.
My current manual process is roughly as follows.
For text:
-
Dust & Scratch in Topaz removes small dirt and defects well, but it also smooths the text and sometimes destroys dots, diacritics, and parts of letters. Frequency separation can partly solve this, which again shows how useful a universal frequency control would be.
-
In Photoshop, I create a copy and apply Dust & Scratches so that the text disappears while the background remains. I then use that result in Divide blend mode. This removes the paper background and uneven lighting extremely well, even from badly photographed pages where ordinary threshold-based methods start behaving poorly.
For photographs:
- I use a descreening plugin to remove the printed halftone pattern, and Denoise Max when necessary.
Combined:
- After that, I use Wonder 2 and then Wonder 3. For some reason, Wonder 2 often restores the shape of the text better, while an additional 1x pass through Wonder 3 makes the result less smoothed.
I would like to replace this entire process with one model. It should remove dust, dirt, paper texture, uneven lighting, printed halftone patterns, and scanning defects; restore text without smoothing the letterforms; and then upscale the result well.
Fully automatic correction of complex page geometry would also be useful. I do not mean only straightening a rectangular sheet, but correcting page curvature and complex perspective distortion.
I would provide two modes. The first would be for compressed JPEGs, photographed pages, and poor scans. The second would be for high-quality source scans and should work as conservatively as possible.
For good scans, the application could also recommend an appropriate input resolution. In practice, 600 dpi often provides excellent source material for text restoration, while excessive resolution starts capturing the texture of the paper, the printing process, and other defects that later have to be removed.
- Strange hot pixel correction
I have encountered cases where a camera had a fairly worn sensor and a large number of defective pixels. In my experience, the current correction behaves more like the Dust & Scratches filter in Photoshop: it removes some bright elements while still failing to eliminate all the bad pixels. Also, this function should be moved directly into RAW Denoise rather than hidden in the general settings.
- Smart Downsampling
The user presses Auto, and the program detects false resolution and downsamples the image only as much as necessary to remove empty, blurry pixels while preserving real detail for later upscaling. This should not be a simple Bicubic Resample, but a more advanced algorithm that optimizes the image for further processing.
Something similar already existed to some extent in Recover and is now used in Wonder 2, but it should also be available as a separate tool.
An additional option could optimize the image for later compression. The algorithm would redistribute detail and contrast in advance so that important edges and textures survive JPEG compression better, while the reduced version retains as much useful information as possible for later restoration and upscaling.
- Preview Exposure for RAW Denoise
Sometimes a RAW file is so dark that it is impossible to judge the noise or the effect of denoising properly. Adjust Lighting is too slow for this purpose and changes the actual processed result.
A simple temporary Preview Exposure slider is needed. It should brighten only the preview and have no effect whatsoever on the exported image.
- Proper model comparison
At the moment, if I have eight Sharpen models, I have to choose one, wait for it to render, remember the result, switch to the next one, and wait again.
There should be a Model Comparison Grid for viewing several versions of the same area at once, plus Pairwise Comparison, where variant A remains fixed while it is compared sequentially with variants B, C, D, and so on.
- Add a True OLED dark theme
I would like a separate theme with genuinely black main panels. The background around the image itself should remain configurable and neutral.
- Full interface and model customization
Users should be able to hide any models they do not need, delete them from disk, and prevent them from being downloaded again automatically.
It would also be useful to customize the placement of major interface elements. For example, I find the button that sends the result back to Photoshop less convenient at the top than it would be at the bottom next to Export. That is subjective, so the better solution is simply to let the user choose.
The models and enhancements could also be organized more clearly. Super Focus v3 and Sharpen logically belong near each other, and Wonder and Upscale are also closely related. Denoise Max, after all, already lives inside the Denoise menu.
Another option would be to divide models into generative, conventional quality-enhancement, and artistic categories.
Users should also be able to drag models around, add them to Favorites, separate or group them as they wish. For example, someone could move Denoise Max directly into the main menu while placing Wonder inside the Upscale menu.
