Hey there, I'm Sats. This is a sideblog. Please tag nsfw/fetish stuff.
Inflation hit like a truck harder than we were expecting during winter, so i’ll be opening emergency commissions for the foreseeable future to help make ends meet going forward. Any and all help would be appreciated.
I’ve recently mentioned that I would like to write more about some of the techniques I use, and this last search presented the perfect opportunity.
Here, we see that if we pass the starting image to SauceNAO as is, all that it returns are repeated results, seemingly identical to the edit we have. However, with a simple tweak, we can change that.
We start by cutting out all sections of the starting image that do not appear to belong in the original—in this case, that’s only the overlaying text. It’s better to keep as much of the underlying image as possible.
Next, we fill these holes with something that looks like it belongs in the image. Since I have Photoshop, I use the Content-Aware Fill tool, but you could achieve a similar effect by copying pieces from the remaining sections.
Lastly, we apply a soft Gaussian Blur over the filled sections to reinforce the uniformity over the image. If you squint and it looks like the image is consistent, it’s working.
Now, if we pass the modified image to SauceNAO… et voilá! A result for the original image with proper attribution shows up. Isn’t that cool?
I can try to explain the reasoning behind this technique, but since I don’t know the inner workings of their system, think of this as informed speculation only.
There are many ways to build a Reverse Image Search system, but that will often involve the construction of a descriptor—something that captures the visual characteristics of an image—traditionally created out of a downscaled version of the input. That might be due to the complexity of the construction process or to size requirements for storage.
Above, I shrunk the images we used into 32x32 pixel kernels. Notice how, in the starting image, the overlaying letters are still quite noticeable, even in this small size. The modified sample, however, looks strikingly similar to the kernel of the original image, even though it’s just a reconstruction without any additional information.
That’s my guess, at least. Now, this might not work that often, particularly when most of the underlying image is covered, but I hope that it can be useful to some of you!
Chainmail for rabbits in iron, sterling silver, jute fabric, and canvas fabric. One of my master’s theses has since also been shown at the National Museum’s exhibition “Konceptdesign”. - Emma Billbäck
