[Paul the Sunman]

ON1 Resize AI. Coming in April, integrated into ON1 Photo RAW 2022 (free for owners and subscribers) and as a standalone/plugin.

Existing methods of photo enlargement, or interpolation, make up new pixels or resolution by copying those around them, which often leads to enlarged photos with less detail, soft edges, and more noise. ON1 Resize AI uses deep machine learning to study different details we see in the world, whether natural textures like tree bark, animal fur, feathers, grass, leaves, rock, and skin to man-made textures like concrete, fabric, or brick. Studying millions of examples of the everyday world will allow the software to upscale photos and their textures to virtually any size, leaving photographers with ultra-high resolution images.


I was wary of their hype with NoNoise AI, but it turned out to be even better than they said, so I'm hoping for the best on this one too. Who meeds a 60 Mpxl camera?

[biz-engineer]

Who meeds a 60 Mpxl camera?

It's not so much the megapixels, but false colors due to insufficient pixel density. AI doesn't play well in presence of color aliasing (my experience with up-sampling, AI for printing).

For prints, I would rank image imperfections by order of importance:
- lens CA (striking, immediately visible in prints)
- aliasing, false colors, moiré (catches attention before lack of resolution becomes visible)
- edge and corner sharpness gradient from center (distracting, depending on optical properties)

Pure pixel level resolution / detail doesn't play a big role because the eye sensitivity to contrast isn't constant with feature size. For example at 20x30" print size, it is difficult to tell the difference between a 200 PPI sharpened print and a 300 PPI with less sharpening, the eye get fooled by very small details. Pixel shift eliminates color aliasing and moiré pattern, which brings a significant improvement for large prints. Unfortunately, pixel-shift doesn't work for all shooting situation, higher resolution sensor eliminates false colors in all shooting conditions.

[normhead]

It's not so much the megapixels, but false colors due to insufficient pixel density. AI doesn't play well in presence of color aliasing (my experience with up-sampling, AI for printing).

For prints, I would rank image imperfections by order of importance:
- lens CA (striking, immediately visible in prints)
- aliasing, false colors, moiré (catches attention before lack of resolution becomes visible)
- edge and corner sharpness gradient from center (distracting, depending on optical properties)

Pure pixel level resolution / detail doesn't play a big role because the eye sensitivity to contrast isn't constant with feature size. For example at 20x30" print size, it is difficult to tell the difference between a 200 PPI sharpened print and a 300 PPI with less sharpening, the eye get fooled by very small details. Pixel shift eliminates color aliasing and moiré pattern, which brings a significant improvement for large prints. Unfortunately, pixel-shift doesn't work for all shooting situation, higher resolution sensor eliminates false colors in all shooting conditions.

I'm largely in agreement. All but the last one. Higher resolution eliminates false colour until moire and colour bending set in, then high resolution causes false colour and moire.

I've used the Topaz Gigaixel AI with mixed results. Sometimes it works well, sometimes I just toss the images.

[BigMackCam]

It's not so much the megapixels, but false colors due to insufficient pixel density. AI doesn't play well in presence of color aliasing (my experience with up-sampling, AI for printing).

For prints, I would rank image imperfections by order of importance:
- lens CA (striking, immediately visible in prints)
- aliasing, false colors, moiré (catches attention before lack of resolution becomes visible)
- edge and corner sharpness gradient from center (distracting, depending on optical properties)

Pure pixel level resolution / detail doesn't play a big role because the eye sensitivity to contrast isn't constant with feature size. For example at 20x30" print size, it is difficult to tell the difference between a 200 PPI sharpened print and a 300 PPI with less sharpening, the eye get fooled by very small details. Pixel shift eliminates color aliasing and moiré pattern, which brings a significant improvement for large prints. Unfortunately, pixel-shift doesn't work for all shooting situation, higher resolution sensor eliminates false colors in all shooting conditions.

Out of curiosity, do the various AI-based post-processing programs work on TIFFs as well as raw? If they do, demosaicing the raw in something like RawTherapee with the most suitable algorithm (with "false colour supression steps" set accordingly) first, exporting to 16-bit TIFF, scaling to the required size and then using the AI software on that TIFF, might work better. In my head, that thought process makes sense... Might be worth trying?

[jlstrawman]

Out of curiosity, do the various AI-based post-processing programs work on TIFFs as well as raw? If they do, demosaicing the raw in something like RawTherapee with the most suitable algorithm (with "false colour supression steps" set accordingly) first, exporting to 16-bit TIFF, scaling to the required size and then using the AI software on that TIFF, might work better. In my head, that thought process makes sense... Might be worth trying?

Mike,

From The ON1 Resize User Guide:

"ON1 Resize 2022 can view and open the following file types, in RGB color, at either 8/16 bits per pixel:

• Raw files from most major camera manufacturers (file extension varies by manufacturer)
• Digital Negative (DNG)
• Photoshop (PSD)
• Photoshop Large Document (PSB)
• Tagged Image File Format (TIFF)
• Joint Photographic Experts Group (JPEG)
• Portable Network Graphic (PNG)
• High Efficiency Image Format (HEIC)

When you wish to export or share images, Resize can save files in PSD, PSB, TIFF, JPEG, and PNG format."

[Paul the Sunman]

Mike,

From The ON1 Resize User Guide:

"ON1 Resize 2022 can view and open the following file types, in RGB color, at either 8/16 bits per pixel:

I think that's for the current Resize. No guarantees for the new Resize AI, which I think is a completely new product.

[normhead]

Out of curiosity, do the various AI-based post-processing programs work on TIFFs as well as raw? If they do, demosaicing the raw in something like RawTherapee with the most suitable algorithm (with "false colour supression steps" set accordingly) first, exporting to 16-bit TIFF, scaling to the required size and then using the AI software on that TIFF, might work better. In my head, that thought process makes sense... Might be worth trying?

In my workflow, I edit a Photo, post process do all adjustments except noise reduction and sharpening. Then if necessary Denoise with Topaz Denoise AI, and Sharpen with Sharpen AI. The whole package GigapixelAI, DenoiseAI and Sharpen AI were on sale for $99 just before Christmas. Even cheaper if you got the Forum Discount.

I find even shots that are pretty good look better with some sharpening, in this case her face was slightly out of focus. The default is 30, if my original images is good, I may used value as low as 10, so I don't introduce artifacts.


It does more than sharpen, it accentuates lower contrast images. It's actually a digital manipulation of the image to produce a more pleasing result. I'd guess my images are probably 80% sharpened at the moment. I Denoise and sharpen at full size (from my K-3 192 MB per image) then export as jpegs to 3840 x 2160 for my TV and the forum as 1MB jpegs. I also save the full size denoised sharpened tiff to a drive in case I ever want to print.

---

Last edited by normhead; 03-15-2022 at 02:49 PM.

[BigMackCam]

Mike,

From The ON1 Resize User Guide:

"ON1 Resize 2022 can view and open the following file types, in RGB color, at either 8/16 bits per pixel:

• Raw files from most major camera manufacturers (file extension varies by manufacturer)
• Digital Negative (DNG)
• Photoshop (PSD)
• Photoshop Large Document (PSB)
• Tagged Image File Format (TIFF)
• Joint Photographic Experts Group (JPEG)
• Portable Network Graphic (PNG)
• High Efficiency Image Format (HEIC)

When you wish to export or share images, Resize can save files in PSD, PSB, TIFF, JPEG, and PNG format."

Thanks, Jerry

So, to @biz-engineer - per my previous post, I think using the most appropriate demosaicing algorithm (with false colour suppression steps) to process the raw file, exporting to TIFF and then using resize software might yield better results...

[BigMackCam]

In my workflow, I edit a Photo, post process do all adjustments except noise reduction and sharpening. Then if necessary Denoise with Topaz Denoise AI, and Sharpen with Sharpen AI. The whole package GigapixelAI, DenoiseAI and Sharpen AI were on sale for $99 just before Christmas. Even cheaper if you got the Forum Discount.

I find even shots that are pretty good look better with some sharpening, in this case her face was slightly out of focus. The default is 30, if my original images is good, I may used value as low as 10, so I don't introduce artifacts.

Thinking about the properties of images and the processing pipeline in most raw conversion software, the most effective approach should - I believe - be to demosaic according to the image and intended use case (e.g.. for a high detai base ISO image, using the AMaZE algorithm - or, for a high ISO, high noise image, LMMSE), then apply colour (but NOT luminance) noise reduction including any hot pixel or line artefact filtering, carry out any general processing, then move onto regular or AI-based noise-reduction and/or resizing.

It does more than sharpen, it accentuates lower contrast images. It's actually a digital manipulation of the image to produce a more pleasing result. I'd guess my images are probably 80% sharpened at the moment. I Denoise and sharpen at full size (from my K-3 192 MB per image) then export as jpegs to 3840 x 2160 for my TV and the forum as 1MB jpegs. I also save the full size denoised sharpened tiff to a drive in case I ever want to print.

One element of - or approach to - sharpening is to increase local contrast, and in fact one of the reasons we perceive some lenses to sharper than others - even if, on the test-bench, they actually resolve similar levels of detail - is due to so-called "micro-contrast"...

[RobA_Oz]

Thanks, Jerry

So, to @biz-engineer - per my previous post, I think using the most appropriate demosaicing algorithm (with false colour suppression steps) to process the raw file, exporting to TIFF and then using resize software might yield better results...

My workflow is to open in PS Camera Raw, perform any necessary adjustments there (such as reducing highlight blowouts, straightening etc) before opening in PS and then using Sharpen or Denoise AI as a plug-in filter, which automatically saves as a TIFF. I sometimes use Topaz Studio (the free version, again as a plug-in to PS) because of its range of amenities.

[Paul the Sunman]

I do everything in ON1. Not keen on jumping from program to program.

[biz-engineer]

Out of curiosity, do the various AI-based post-processing programs work on TIFFs as well as raw?

Topaz AI works on raws but there isn't any tool to edit the raws.

[RobA_Oz]

I do everything in ON1. Not keen on jumping from program to program.

Plug-ins make that easy, and I find that no single program does everything as well as others. I use ON1 on my 2020-model MacBook Pro, but my main processing is done on my 2010 iMac, however I recognise that ON1 is offering some more AI filters now.

[BigMackCam]

Topaz AI works on raws but there isn't any tool to edit the raws.

I just read the following article:

https://capturetheatlas.com/topaz-gigapixel-ai/

It's a couple of years old now, so maybe the advice has changed, but it seems like Sharpen AI and/or Gigapixel AI can (indeed, should) be applied later in the workflow after raw conversion and editing. 16-bit TIFF would seem to be a good choice for that. So, you could - I'm quite sure - optimise demosaicing and related false colour suppression at the raw conversion stage before using AI re-scaling...

[biz-engineer]

So, you could - I'm quite sure - optimise demosaicing and related false colour suppression at the raw conversion stage before using AI re-scaling...

Yes, that's what I do. Although I see a significant improvement when up-sizing from pixel shift files. I once made a post here claiming that Pentax default image processing isn't good enough to capture all the information / potential contained in pixel shift files.