Heyo, back posting some behind the scenes stuff here after awhile.
Teases and that sort of thing have generally gone to our Patreon members in the WtW Discord, and some of them know I've been working on this.
I've been working on a number of processes to do advanced denoising of scans in the future, since the handful of better denoisers are:
A- Expensive
B- Time consuming.
I've dialed in a few different things that let me come up with my own solution to potentially use on future scans, and thought I'd let people preview it, and give their thoughts.
I would recommend clicking each image so you can see it full-size and properly compare.
First up, an ancient scan. This is a scan from shortly after I first took over WtW on the left.
The differences here are rather striking, with the denoising blending the colors in a really impressive way, and at a glance, it appears no actual detail has been lost.
I took the jpg, did the new process to it, and ended up with the right image. I then jpg compressed that, and put them both on a lossless PNG for you to compare.
In fact, all of the images were done that way. They were all done off 'final' already compressed images, which is kind of a 'worst case' scenario for doing this kind of work.
This one is from the tri. Part 5 BD booklet. The scans on this gave me problems in general due to having a new replacement scanner at the time.
The results look pretty good, although a decent chunk of 'noise' in areas like her hair ended up with a slightly smudgy almost 'pudding' look to it. This look is generally always there when something is denoised and the color isn't super consistent, it's just a matter of how noticeable it is, and on Hikari it's noticeable.
This is likely a side effect of how 'large' I put scans up, and without simply shrinking scans, would likely hit this type of wall.
It's also possible that's a side effect of the paper, as if you look carefully, you can actually see a similiar effect under the noise on the top part of the comparison.
Kouichi from the new Frontier Blu-ray box. This feels like a great example of the noise on the scan being smoothed out very well.
Haru and his Applidrive Duo ends up being a bit more questionable.
The middle is the original finished scan, the right is using the same level of processing as the other comparisons, while the left is me pulling back a touch.
Likely due to the the actual quality of the print, Haru when he's cleaned up appears a bit overly artificial and smudgy. A few places in his hair where the lines weren't necessarily as defined in the print as they could have been come out a bit weird looking. They were weird looking in the original scan, but the high level of noise ends up hiding it.
The CGI Applidrive tends to have the same issue, the actual linework on the scanned image wasn't always defined well and the noise covered up issues. As an example, the 'G' in Gatchmon in the original scan isn't actually fully readable as a G due to some of the visual effects on the nameplate, but the noise muddies that a bit, and once it's cleaned up, it becomes more obvious.
Overall it's probably improved with the noise removal, but that does add in some things.
Another one that is slightly more questionable...
An old card scan of Lowemon.
Overall it worked well in removing the noise, but in various places, especially in regards to card elements like the various pattern overlays, which look more random in general, it feels like their texture is reduced or missing. Some of the text also has a smudgy and muddy quality to it.
This is much lower resolution and done with different processes than we have now (the actual scan was done by rian as I recall, using the processes the site used for scanning at the time), which may account for some oddities, and may not even apply if I had a card on hand to scan and test with.
Last comparison for now...
This last one I've used before, but I updated it.
This is one of the raw scans from tri. Part 3, shrunk to roughly the size I finalize edits at.
The first is the raw.
The second is after descreening.
The third was the advanced denoiser I had a trial for a few years ago, as mentioned above. Denoising is always done 'after' descreening.
The fourth is using the new process to denoise.
The new one compared very favorably with the old test, it looks less blurred overall, and the purposeful 'smudgy background' is intact and looks good. A few places it appears to be less blended than in the prior test (most notably there are some visible diagonal lines in the grey block to the right of Patamon's wing.)
The process seemed to make one bit that wasn't cleaned up in the original scan far more noticeable though, a little dot a bit above where Patamon's leg connects to his body. The denoising made that stand out so much, I thought my monitor had a small crack on it in fact.
The processes, software, etc. used on various site elements are always in flux, with us looking for improvements, changes, etc.
I'm leaning heavily to using this process (or at least variants of it) on future scans, and any input or thoughts would be appreciated.
Supporting us via our
Patreon or
donations always helps if you can.