Only now noticing Denoise MAX and Wonder have a tendency to round of the dots on any text, even ones that have square dots.
EDIT: This is at lower DPI/res, it doesn’t do it if the DPI is high enough. And Text preservation does not really help with it either.
Thanks for the feedback! What you’re seeing in these low-resolution images may not actually be noise, but compression artifacts. In that case, another model may be better suited for the image, such as Wonder.
If you’re still getting unwanted results, you can send the original image along with the unwanted result to support@topazlabs.com. We can test the file on our end and, if needed, relay it to the development team as a user case where the current models aren’t giving the expected results.
There is no compression going on, these are scanned at high resolution, then downsized, saved as PNG. ( Mostly trying to find a way to get better denoising, at the higher dpi ti doesnt work as well, the problem here is mostly the low resolution obviously makes it impossible for the models to know what font is used.)
@drdre1411 - the stepping seen is not noise, it is compression. You can test different Upscaling models, meant to fix compression or Wonder, which is also an upscaling model! Noise is much different, and is present in low light photography or photos shot in RAW format. It is not noise that we are seeing here! The stepping caused by compression is most likely happening in your downsize step and PNG export. You can rescan the files true to size, it will give better results.
Its not compression its the interpolation from downsizing and descreening. The png is saved at highest setting in photoshop.
The stepping you’re seeing is aliasing, and this can be addressed with the Upscale models, which are designed to resolve these types of artifacts. There are multiple Upscale models you can test, as well as Wonder. This is not noise, so the Denoise models would not be the right enhancements to fix it.
If the original high-resolution scan does not have this stepping, the best approach would first be to identify which step in your workflow is introducing them and remove that step, rather than trying to fix artifacts that weren’t present in the original image.
If you’re downsizing the image in Photoshop, for example, that may be where the aliasing is being introduced. I would recommend doing the resizing in Topaz instead and seeing if that avoids the issue altogether.
If that step needs to remain part of your Photoshop workflow, then you can use the Upscale models in Topaz afterward to help correct the introduced aliasing. There are multiple Upscale models you can test, as well as Wonder.
But ideally, preventing the jagged edges from being introduced in the first place will give you a better result than trying to recover them afterward.
Let us know how it goes!
I think there is a misunderstanding here, its not the aliasing that is the problem, its the fact the upscaler ( wonder does it too yes) and max denoise turn square/reactangular dots on the i/j and periods at the end of sentences into rounded ones. Even with MAX Denoise I have no issues with how it handles text otherwise.
I see! I’ve added this as a Feature Request for the ability to preserve this type of detail when upscaling, such as keeping the period square rather than having the model interpret it as aliasing and turn it into a rounded dot.
You can add your vote to the request as well. The development team reviews the most popular requests when considering future implementations.
More broadly, the request would be for a way to tell the Upscale models to preserve certain characteristics of text rather than reconstructing them differently.
Both Wonder and Denoise Max are generative models, so they can reconstruct details rather than strictly preserve the original pixels. Wonder could be a good model to test for this particular case. Denoise Max would not really be the correct enhancement for this issue, since the problem isn’t noise. You can still use it if you find that it helps your workflow.
Thanks for helping clarify exactly what you’d like the model to preserve!
