[TEXT] - "Preserve Text" AI Model that refines text to be like v2.4.2 or Gigapixel's "Text & Shapes"

Hello. I think a great addition would be text improvement. Similar to Face AI, but geared towards improving text on images. It usually happens that low resolution text turns out jagged, or even further mangled by the normal processing - sometimes the artifacts are more subtle, like swirls of colour inside a solid or gradient fill text. This would also set Photo AI further apart from the competition

One of my favourite suggestions. I recently scaled up a video from 720 wide to 1280 wide the old way before processing to avoid the text to get messed up! also happens with video ai.

This is in active development and an initial version is currently in the latest beta release. If you would like to try it out, you can apply to be included in the beta testing group here. We are actively seeking any feedback on this feature.

I’d put a +1 here for this. I shoot airshows and planes a lot, and if I use sharpen then the text on planes, their serial numbers etc often end up looking odd.

Just requested to be added! Thank you!

Hi, The Topaz team are working on it I’ve included an except from the Roadmap page

Feature Roadmap

We’re now actively working on:

  • Adding a Refine Text filter that applies a mild deblur and removes artifacts from detected text.

Text is not able to be improved at all. The original image is better than the Preserve Text or any other option.

I was trying to enhance an image so I could read some text in an image and while Topaz works great for other things, text did not work at all. It would be a very nice feature.

We are actually working on a text recovery model! It’s in Beta right now, but keep an eye out for it in a future release.



I use Topaz AI for race cars. There are lots of small sponsor stickers with small text. The small text ends up looking like letters from the Greek alphabet after processing. I agree it would be a nice improvement if text processing/recovery would be added to the software.


I’ll second this. I use Photo AI for action shots of motorbikes. It works well for the overall picture, but the text and logos on the bikes, hemets, jackets, etc comes out very rough and almost metallic looking.

The text recovery model metioned by @preston.werner sounds like it could be what we need! Can’t wait! :slight_smile:

1 Like

It would be great for Enhance if there was automatic text recognition.

This would make the work of mass image processing much easier.

Unfortunately, the AI models have no understanding of text.

And if it were trained with text, there would always be the possibility that it would create letters somewhere instead of normal pixels.

Thanks for the idea and suggestion! Make sure that you do not forget to vote for your idea :slight_smile:

The text & shapes model added to Gigapixel AI is extremely good.

Would it be possible to use it for when we mask a zone with text in Photo AI?

Or simple shapes.

It could work as it does now, with just a checkbox or setting to activate, and the text & shapes model would be used just for the selected part of the image.

Photo AI - Preserve Text

Gigapixel - Text & Shapes

The second one is more pleasing. My original image is a pretty extreme example, but I really think this combo could give less alien results for small text.

Here’s the original tiny image. That’s an extreme example, but it shows how it would be even better with slightly larger base images

Just some other examples :
With Preserve text

Text and shapes model


Preserve text

Text and shapes model

I really think having a specialized model just for certain parts of the images, like texts and shapes, woulb the way to go. Preserve text could have this option.

No one model can be good at everything. And no model has to be perfect. I personally never used the model Graphics.

A new model brush could be a good idea too, maybe.

So we could have specialized models for various generic elements and paint certain objects that

Realistically, we can’t have a model for every single object or concept, but text, signs, clothes, nature, cities, and roads, for example, cover a large percentage of what’s in pictures in general.

  1. Very good generalist models like those we have today. The high fidelity, the low resolution, etc.
  2. Very good specialized models trained and more specific concepts that can be used just for some parts of the images. And could be used alone too if we have the object or concept alone in an image.
  3. These could be used on the upscale image too. A sign or text or whatever is good but not great after the first pass, just brush over it, select the model related to what you want to enhance further, and voilà. Less need to do more compositing in other softwares

Hello @roberto.avanzi and everyone!

We have released today in our v3.1.0 a major improvement on our Preserve Text enhancement, and would like you to test it and let us know if you have any feedback :slight_smile:

You will be able to download this update from your app directly or by using the following link here:

Here is the full list of changes/fixes done on this release. Please let us know if you have any issues in updating to this version!