Tale of a Lost 300D Photo and AI Uprezzing

These are very nicely done - thank you for sharing. 

Henry Richardson wrote:

All the recent talk about Adobe Super Resolution, Gigapixel AI, etc. has got me to thinking about a May 2005 photo that I lost that has always given me some pain. Since January 2000 when I got my first digicam the only photos I have lost are all the ones I took in May 2005. It was only about 50 and I didn't even discover that I had lost them until about a year later when I noticed that my carefully created folder of OOC photos that contained subfolders for the year and month was missing the May 2005 folder. I recall that I had been having problems with my computer in May 2005 and after trying everything I could think of to fix it, searching on the internet, asking about it on a PC forum, etc. I finally decided to do a clean install of WinXP. I very carefully (I thought!) ensured everything was backed up, had all my software CDs to reinstall my programs, etc. I then did the clean install, but it didn't fix the problem. Apparently my careful backup somehow missed the May 2005 photo folder though. After I discovered the loss the sting was even greater because the clean install of WinXP had all been for nothing.

I generally take the photos I like the best, prepare them, and put them on my photo website. I looked at those photos and discovered that there was only one photo from May 2005 that I had judged was worth putting up so I felt relieved that I apparently didn't lose a bunch of good photos. The trouble was that the one photo is a photo I like a lot. All I have is a 640x507 pixel jpeg version of that photo that I took with my 6.3mp Canon 300D that made 3072x2048 pixel files. I had apparently cropped it a bit, possibly made some small adjustments, and used a bit of unsharp mask before resizing, and saving a medium quality jpeg for the website. In those days I was using Paint Shop Pro X. The small jpeg I have has no EXIF since PSP X would strip it off when you saved. Basically this is almost a worst case scenario for trying to sort of save a photo.

Yesterday I tried taking that tiny 640x507 pixel jpeg and doing the following:

1. Photoshop Preserve Details 2.0 to increase the size to 2560x2028 pixels in one go.
2. Photoshop Preserve Details 2.0 to increase the size to 2560x2028 pixels in 2 steps: 1280x1014 and then 2560x2028.
3. Photoshop Super Resolution to increase the size to 2560x2028 pixels in 2 steps: 1280x1014 and then 2560x2028.
4. Photoshop Super Resolution and Preserve Details 2.0 to increase the size to 2560x2028 pixels in 2 steps: SR 1280x1014 and then PD2.0 2560x2028.
5. Photoshop Preserve Details to increase the size to 2560x2028 pixels in one go.
6. Photoshop Bicubic Smoother to increase the size to 2560x2028 pixels in one go.

I judged that 1, 2, 3, and 4 were pretty similar and that 5 and 6 were clearly worse. I decided that 1 was probably the best though.

Last night while searching for some more info I saw someone saying that some of the online AI photo enlargers do a better job than Photoshop Super Resolution so this morning I used one of them to make the photo 2560x2028 pixels. I decided that for this particular photo it was better. I think it would probably make an okay 8x10 even -- maybe -- as long as you aren't viewing it at an unnaturally close distance. I think I would add some grain in Lightroom to give a bit more texture though. The uprezzed files all have a bit too much smoothness to them. I imported the new 2560x2028 file into Lightroom and here it is at 800x634. He is just hanging around in Vancouver, British Columbia:

Canon 300D + Sigma 18-125mm f3.5-5.6

My understanding, which may be wrong, is that resizing algorithms such as bicubic, Lanczos, etc. try to interpolate to a larger size by keeping the same details as in the original, but making everything bigger and trying to maintain smoothness without pixelation. Most will remember Genuine Fractals that was popular around 20 years ago when we had 1-5mp digital cameras. My understanding is that it tried to actually create new details as it made the photo larger to make it look more natural. I recall that, I think, Nikon at a trade show had a big poster (1-2 meters) made from one of their 3mp digicams (Nikon Coolpix 990) that had been resized using Genuine Fractals. Maybe it was an eagle or something else. Anyway, it impressed many people and there was lots of talk on dpreview at the time. Now these new AI algorithms apparently do this even better.

Personally, I don't have much use for this stuff to make my 20mp and 16mp photos even bigger. But thinking about it some yesterday I remembered this old 2005 photo that I felt such regret that I had lost the original. I also have some 1.3mp Olympus D-340R (bought in January 2000) photos that my wife took using the 640x480 mode -- despite me always telling her to only use the full 1280x960 mode -- old photos of family and friends that are just 640x480. I plan to see what the AI resizing can do for some of them. Also, some old photos made with 2-6mp cameras that I had cropped might benefit.