Initial down-sample for digital photos of prints

Sorry if this was previously suggested – this is my first time posting, and I didn’t immediately see anything related in the topic list…

I have a number of (digital) images of older family fotos, which I took using an iPhone from the printed photos. This may be too much of an ask, but after playing with the digital images around for a bit, I found that by first applying a down-scaling to the images and then using Photo AI, it did a better job than by trying to use the image that I already have, possibly because there is so much additional noise in the image that throws the AI models off.

For that purpose, I was wondering if (in an ideal world) Photo AI could detect from the loaded image what the “optimal source resolution is” (as part of the autopilot), and offer that as an initial/intermediary step prior to upscaling and recovery?

Great idea - we’ve heard this before, although I’m not sure if it’s in the Ideas forum. Would you mind sharing one of the images that would benefit from this? Can upload to our Dropbox if you don’t want to post it to the forum.

Hey Eric,

Thanks for getting back! I’ve uploaded a ZIP file with one image where I played around with. Depending on the “object” (or subject) in the image, different resolutions seem to work best.

I have to admit I do not know exactly what the source of the “interlacing” (like) artifact in the digital photo is – this one was taken quite a few years ago (with an older iPhone). Ultimately, my thinking goes in the direction of having a variety of “plugins” available that people can flexibly put into a “preprocessing pipeline” (including down-sampling, and other filter operations, which could be selective based on some kind of pattern recognition, to avoid having to do many passes with the same photo and later stitching together a final result manually).

And I appreciate that this might go well beyond what you guys plan for the main product… maybe somewhat similar to the idea I added to the other thread (Topaz offering a two-pass video editing with face-identity recognition and a LoRA-like additive model training in the cloud), I personally would also be happy to pay some extra bucks for plugins – especially if they substantially address niche demand that most customers wouldn’t be interested in.

As an example, about 35 years ago, I appeared in a school theater production, and one of my uncles recorded this onto VHS tape. My oldest brother at some point digitized this, and he will send me a DVD image this upcoming weekend, and I will play around with Video AI, but I will likely also try extracting a few frames (those with as little noise as possible) to see whether the models available in Gigapixel and Photo AI might show different/better results. It would then be fantastic if (should those results be better) I could “pipe” frames (of certain scenes, say) easily through the pipeline in Photo AI – although I have already had the experience that Video AI probably has much better support for not letting MPEG compression artifacts get in the way of good upscaling quality :wink:

Anyway, so far I love your products! :slight_smile: Cheers!!

Thanks for sending the images over! I had a look for comparison and it seems that the downscaled image did not have remove noise turned on.

Remove noise is strong by default and is removing fine details from the image by treating them as noise.

Downscaling the image increase the quality and reduces the noise enough that remove noise is not needed.

I shared this with my team so we can look into improving the remove noise model and figuring out which file sizes Topaz Photo AI currently works best with.

We may need an improved remove noise model that does keeps the details better without removing them.


Hey Lingyu,

Thanks so much for getting back! Yeah, I think that if the remove noise model could be adaptive to work for different types/sources of noise (at different scales), which it auto-detects from the image, that would definitely make a manual downscaling unnecessary!

1 Like

Could you try using the new Remove Noise beta models and see if that is working better for you?

In the new update to Topaz Photo AI we are using the new models by default which should retain details better.

Hello Lingyu,

Thanks for letting me know there is a new beta model! :slight_smile:

I’ll try it tonight when I get home! Will the regular product version allow me to install the beta model, or do I need to download it through a special link? – I’m using an M2 MacBook for testing so far (I assume that the model files are platform independent, but just wanted to mention it, in case…)

Sorry I didn’t get to it last night, @Lingyu

Here one example image:

Followed by a manually downsampled version:


And then the two versions upsampled with the default settings in v1.3.7 (with Normal vs Noise Removal model):

While many “details” are missing in the second version (and the cat definitely looks very “unnatural”), by first downsampling manually, it still created a somewhat cleaner overall picture…

If you have any ideas for how to improve on the settings of the program, that would be greatly appreciated!! :slight_smile:

Thanks for running the test. It looks like the downsampled version is still the best, but the Remove Noise v2 is doing better.

It would be best to decrease the Strength as much as possible to retain the most detail. It still looks like the effect is too strong.

That would be the best advice I have at the moment.