Image size for AI

[Festive_epicness157F]

The Adobe guidelines says " ... meet the Adobe Stock minimum file size requirement of 4 megapixels. Use the Enhance tool in Lightroom to increase the original file size, but don’t go beyond 6 to 8 megapixels."

But isn't that a rather small image ..? I believe that images usually are almost 20 megapixels - e.g. 5.000 x 3.500 pixels? Do I misunderstand something?

[Nancy OShea]

See these links for more details:

• https://helpx.adobe.com/stock/contributor/help/quality-and-technical-issues.html
• https://helpx.adobe.com/stock/how-to/tips-stock-image-acceptance.html
• https://helpx.adobe.com/stock/contributor/help/generative-ai-content.html

[Festive_epicness157F]

Thanks for the usefull links, Nancy. But I was only interested to find out if 6 to 8 megapixels are not very small images.

[Nancy OShea]

Most modern cameras shoot in 24-50 megapixel range, far more than what AI services can generate.

You should NOT upscale a tiny image beyond a certain threshold.  2x linear resolution is about the best you can do with commercial Firefly in Lightroom as an example.  If you go much higher than that, you'll introduce unwanted distortion & artifacts.

[daniellei4510]

I normally upscale AI images up to between 6000 and 9000 pixels on the longest side. It CAN be done using the appropriate AI upscaling tools and following up with some additional sharpening methods in Photoshop (along with CAREFUL inspection of the result). If an image starts falling apart during the process and begins showing signs of softness, over-sharping, artificacts, haloing and other issues that are beyond the scope of being corrected in Photoshop, they are consigned to the trash. I figure it's a rare occasion when someone would require an image that is 9000 pixels wide, but virtually everyone knows how to reduce an image without the need of special software. While if someone purchases a 6 inch image and needs it three times that size, they may not have access to tools or expertise that would allow them to do so successfully.

[Festive_epicness157F]

Hi Danielle. Thanks for a comprehensive answer. I've looked a bit at Midjourney and have found that the upscaling apps that work best also cost some money. 

[daniellei4510]

You might want to do a search on YouTube with the keywords "PIXImperfect image upscaler." He reviews 7 upscaling platforms, some of which I believe are free to use. Not sure.

[Festive_epicness157F]

Thanks Danielle. I have seen the video. Do you use Photoshop?

[daniellei4510]

Yes. I use Neural Filters > Super Zoom most of the time. I have Gigapixel AI but it tends to crash on my old 2015 iMac.

[Nancy OShea]

Midjourney's previous version 4 had a good upscaler that transformed 512 x 512px images to 1024 x 1024px.  But the current version 5 doesn't support upscaling beyond native 1024 x 1024px size.  Rather than go the extra distance to fix the problem, Midjourney told users to seek out 3rd party upscalers.

These are the best I've found:

Topaz Labs Gigapixel AI  - https://www.topazlabs.com/gigapixel-ai

Stable Diffusion's Automatic1111, img2img tab  - https://stable-diffusion-art.com/automatic1111/#Img2img_tab

[Festive_epicness157F]

Thanks Nancy. I found the three first mentioned in Danielle's post very good along with PixelBin. I don't know if it's worth the money to invest in this AI-thing, so maybe mess around with Photoshop a bit. But the small originals are exeptional ... :0)

[daniellei4510]

There's actually an advantage to that for users on a tight budget. Upscaling from 512x512 took a few extra seconds (sometimes another whole half minute or longer) compared to all four images coming in at 1024x1024 with versions 5+. There's really no upscaling involved in that case, since all four images are already enlarged to the max. If you really want to save a few seonds, you can save all four images in the grid at once and separate them in Photoshop. I'm too lazy to do so, but in the rare instance when all 4 results might be worth using, it takes no extra time at all in terms of server costs.

[Festive_epicness157F]

Thanks Danielle - "advantage to that" - to what ... you surely don't mean you just can upscale in Photoshop as we normally do with a small enlargement ..?

[daniellei4510]

No. I differentiate between "enlarging" in Photoshop and upscaling with AI. Super Zoom in Photoshop purports to use AI to upscale and it does a pretty good job. It doesn't have all the controls that Gigapixel AI does, so I'll use that if Super Zoom doesn't do the trick. 

[Festive_epicness157F]

9.000 is extremely big. It's amazing you can go from 1.000 to 9.000 ...!

[daniellei4510]

Yep. 30" at 300 dpi. Works best with portraits (fashion, skin care, children, animals, which I tend to specialize in), where as landscapes, food, wallpapers and the like can start falling apart if I go beyond 20." When it comes to portraits, I hold to the theory that if the eyess are sharp, the entire image will SEEM sharp (within reason, of course). I also rely heavily on facial restoration software, starting with the original image size. Following a 30" upscale, the eyes, lips and nostrils can often become pixelated, so I'll run the faces alone through facial restoration software a second time. I'll then work on one eye to round out the iris, stroke it to sharpen the edges if necessariy, then center the pupils. I'll then copy that iris and paste it over the other iris so that the catchlights are identical (as they would be in the majority of actual photographs). Takes awhile. But at least no one can accuse me for uploading 1000 AI images at a time. Or even twenty for that matter. 🙂

[daniellei4510]

Needless to say, I can't upload the complete image as it would be too large, but here's a sample. The main image is a screenshot of a 9000 pixels wide image taken at Print Size @ 300 dpi (from a 1500 pixels wide original @ 300 dpi). As I mentioned elsewhere, it's highly unlikely anyone would require an image 9000 pixels wide, so it most likely would be reduced for the web or even for print (inset is a 30% reduction). So in the same manner that issues might become more apparent to someone with a very critical eye after upscaling, conversely, they become less apparent after being reduced. At least with respect to viewing an image with the naked eye. I don't doubt there are all kinds of things happening on the pixel level that extreme magnification would make apparent.

[Festive_epicness157F]

Amazing ..!

[Abambo]

The main image is a screenshot of a 9000 pixels wide image taken at Print Size @ 300 dpi (from a 1500 pixels wide original @ 300 dpi).

By @daniellei4510

Very impressive, but there are areas that need corrections:

[daniellei4510]

Not sure what's going on there. Those artifacts do begin to appear when I view the posted image at full size in my browser (but more so on my lower resolution 27" acer compared to my high resolution 27" iMac). but at Print Size on my computer, everything looks normal. The catchlights could still use some detailing, though.

Resolution has always confused me. For example, if I view the image at Print Size and Actual Size, everything looks fine (catchlights excluded). But when I view the image at 100%, the image appears larger on screen and the artifacts begin to appear (in the lips, for example, although not as much as the sample you posted).

Anyway...easy fixes.

Thanks, @Abambo 

[Abambo]

Print size is not a good viewing size. The only size that is accurate on screen to check the quality is 1:1, one pixel on screen is one pixel with the image. All other sizes are interpolated and not accourate. However 200% and above can be used to confirm artefacts, as you simply see them better. 

It may be that I checked the lips at 200%. 

(High resolution screens have smaller pixels. I see artefacts better and faster on my QHD resolution monitor than on my UHD monitor. It's a matter of pixel size.)

[Nancy OShea]

Keep in mind that images embedded on this forum are compressed to save bandwidth which can adversely effect picture quality. 

[Abambo]

Keep in mind that images embedded on this forum are compressed to save bandwidth which can adversely effect picture quality. 

By @Nancy OShea

Those quality issues are no compression artefacts.