WFFJ-TV

By:  David Deschesne
Editor/Publisher,
Fort Fairfield Journal
July 26, 2023

About a month ago, in the June 28, 2023 edition of Fort Fairfield Journal, I wrote an update on the photo-realism that can be achieved with Artificial Intelligence- (A.I.) generated images. The results were stunning, to say the least, but I noted some of the issues extant in those images that made them fall short of perfect photo realism and thus provided clues to determine if the image was a fake, or a real photo. I also noted that “we can expect AI to clean these problems up soon.” Well, soon is now here.

AI image generators were released to the public in December, 2022. The computerized neural networks were trained on the millions of photos posted online and learned very quickly what things are supposed to look like down to the tiniest detail. However, due to the nature at which they generate images, then upscale them, things that have ultra fine detail and nuance—such as skin tones and texture—were rendered a little too plainly. There were very few wrinkles, no pores, no fine hairs and a washed out skin tone overall. To look closely at an A.I. image at the time, the skin resembled that of a toy plastic doll. But, that was a whole month ago which is ancient history at the rate technology is now progressing.

image generated by Stable Diffusion SDXL 0.9, enhanced with Reliberate and upscaled with img2img by Olivio Sarikas. Ref.: https://www.youtube.com/watch?v=Rw7iYrcbGVc


AI generators and upscalers are improving at exponential speeds. Once an image is generated from scratch in one of the top AI generators, such as Midjourney, Stablefusion or Dal-E-2 they can then be imported into other AI image enhancers to tweak and fine-tune them to give them much greater detail, especially to the skin which is really what clues us in to determining photorealism.

In the image above, the original image was generated in Stable Diffusion XL then imported into another AI image enhancer called Reliberate. That app then uses an AI enhancer called LowRa which creates a side light effect, similar to that used by photographers, to use light and shadow to accentuate subtle details. After the additional textures and tones are added the image is then sent to an AI image upscaler to increase the size of the picture to make it look more realistic.

The reason for the upscaler is most AI images are generated in small sizes around 580 x 760 pixels. This smaller size reduces the need for computer processing power so the average home computer can create the image without any special hardware. But, a smaller resolution picture loses detail quickly when it is enlarged to “zoom in” and examine the details more closely. This is due to the nature of the pixel format. When enlarging, all it’s doing is making the pixels larger on the screen. If there aren’t a lot of pixels to start with—such as in a small resolution image—then the image will appear grainy and lose detail when enlarged. This problem is overcome with an image upscaler which adds in new pixels by extrapolating from currently existing neighboring pixels to determine what to create.

This type of upscaling has been done for years but it has its limitations. If you try to upscale a picture too large, the traditional method of extrapolating pixels into the image will cause it to come out grainy and soft because of software limitations not being able to process fine details.

This problem has been overcome with AI enhanced image upscalers. The picture here was upscaled with an AI engine called img2img. What these types of upscalers do is use computer learning to not only add the correct pixels in where they belong to make the image larger, they also use their Artificial Intelligence to decide what features, such as wrinkles, skin tones, etc. need to be added so the end result is a larger picture with the details preserved.

Since the AI is creating all of the details “out of whole cloth” they can be individually adjusted and enhanced by the user by using specialized text prompts to tell the AI enhancer how much or how little of the effect to add.
As you can see in the picture here, the subtle skin tones are preserved along with skin texture and wrinkles. Of course, there are limitations in the presentation of this photo here due to the nature of the printing process. I do use a high end laser printer to publish this newspaper, but the images are not as fine or detailed as they would be if viewed enlarged on a screen or printed via a high resolution photo printer. But, I suspect enough of the detail is coming through here so the skin texture does not resemble that of a plastic toy doll.

The software to do all of this is available at no, or low, cost to the general public. Anyone with a modest home computer can access this technology and create stunningly realistic images. The text prompts to create the initial image are simple English words and phrases like “young, blonde-haired girl in a blue dress walking through a grassy field” so there is no computer code to learn. For the advanced AI image enhancing engines, there is some simple code to learn in order to tell the engine what sub-app you want to use and to what extent you want it to influence the enhancements but those code terms are available online from many sources to teach the user what terms to insert in the text prompt thread and where.

The reason I’m focusing so intently on AI over the past month or so is to educate people how easily they can be fooled now with photo realistic images that perfectly mimic photos taken with a typical digital camera.

The reason this is important is we can never really trust the veracity of a photo again; especially a photo that pops up online from an unknown source. If, for example, you see a photo of a celebrity in a compromising position you will not be able to distinguish if it is real or if it’s fake. If you see a photo of a purported news event, it will be impossible to determine if the photo is real or not.

These are the problems that are going to manifest with AI generated images—a technology that has only been in circulation amongst the general public for seven months at this point.

There is nothing embedded in the meta data of the photo to indicate it is AI enhanced and even if it were, meta data is easily manipulated so it would be useless as a tool to prove authenticity.

Photographic evidence is going to quickly become irrelevant in court proceedings because it will now be so easy to establish a defense that the photo is probably not real due to the advances in A.I.

But what about videos? Video is not as far along in detail enhancements as the photo generators because there is so much more data for the AI engine to process and create. However, video AI engines are progressing rapidly and in perhaps a year or so, videos will also be indistinguishable from reality. In five or ten years, we may be watching Hollywood movies completely generated by AI that are indistinguishable from movies shot with conventional cameras. This means actors and actresses will all be out of work because the characters will be created by the computer with hyper-realism. There will also be no studios, sound stages or traveling to location. Producers won’t have to wait for the right weather conditions when creating outside scenes or fiddling with lighting setups for inside scenes because the entire movie will be generated from the mind of an AI generative art engine via text prompts by a computer programmer. Typical movie goers will not be able to tell the difference between “real” movies and faked AI generated movies.

We are going to have a transformation in what is considered “real” and what is fake. Courts and legislators are going to have to adapt with whole new standards and protocols for the admission of evidence since at that point in the not too distant future, photos and videos as we know them today will simply be inadmissible in court proceedings because it would be impossible to determine if they are in fact real.

There might still be room for news gathering and documentary photos and video of real, actual events. But, online news organizations are already using AI generated images to supplement their news stories because there are no royalty fees such as would be found in stock images taken by actual photographers who license their photos for publication and charge a fee for the use.

Oh, by the way, with AI that whole stock photography market is soon to be dead; nobody needs to buy stock photography for their news story or ad campaign anymore if they can generate exactly what they want with a computer for pennies.

Therefore, it stands to reason that the consumer of those informative media are going to have to exercise a level of discretion most of the general public currently does not possess. It may be that the photographer’s name appended to the photo or video will be used to determine the prima facie designation as a true photo or video image. For example, if you see a photo of a Maine Potato Queen and then my name is credited as the photographer, you can be pretty sure it’s a true photo because it was a public event attended by many people who saw me there taking the picture. That is the level of proof that is soon going to be required to determine if an image is real or fake.

If these photos or videos are used in court, we may have to have the original photographer or videographer there to testify to the veracity of the image entered into evidence and perhaps even show the equipment used to capture the photo/video.

Whatever the solution for determining if a photo or video is real turns out to be, we can be sure that a vast majority of the public will continue to be duped by deep fakes—just look at how gullible people were during the COVID-19 “pandemic” when a vast majority simply believed everything the government and corporate media told them even if it was all easily provable as lies, deception or deliberate manipulation.

Legislators move at the speed of slugs and since they only got the job because they won the popularity contest, it’s unlikely that most of them will even know how to write legislation to deal with these issues.

Obviously, that work has to start now.