​Restoration Is Not Reinvention

---

I have been experimenting with AI-assisted restoration on images taken directly from Charlie’s Angels on Blu-ray, and the more I work with it, the more obvious one thing becomes. There is a big difference between dropping an image into an AI program and telling it to “clean this up” and actually controlling what the program is allowed to do. AI can make an image look better very quickly. That does not necessarily mean it has made the image more accurate. In fact, sometimes the cleaner and more impressive the result looks, the farther it has moved away from the original frame.

I am also not approaching this as someone who suddenly discovered photography. I have more than 45 years of experience taking photographs, plus years working in a wet darkroom, later in Lightroom, and as a graphic artist. I made decisions about exposure, contrast, color, sharpness, printing, and retouching long before any of this technology existed. The tools are obviously very different now, but much of the judgment involved isn't. You still have to look at an image and decide whether you have improved what was already there or simply changed it into something else.

Over time, I have developed a very specific process for working with these Blu-ray frames. I am not going to publish the details of that process because it took a lot of experimentation to develop, and I consider that part of my own work. What matters more to me is the reasoning behind it. I am not trying to see how much freedom I can give AI. I am trying to do the opposite. I want the original frame controlling the result as much as possible, not the software deciding what it thinks the image ought to look like.

That is important because AI is extremely good at filling in things it cannot actually see. Give it a soft area of hair, and it can manufacture individual strands. Give it a face without much fine detail, and it can suddenly produce skin texture, eyelashes, or features that look completely believable. Clothing can become more detailed. Background objects can become more defined. The problem is that believable isn't the same as accurate. The program may create something that makes perfect visual sense without that information ever existing in the original frame.

This is where I think a lot of AI restoration goes wrong. It seems to assume that more detail automatically means a better restoration. I do not agree with that. If part of an image is soft, I would rather leave some of that softness than replace it with details that somebody—or in this case something—has simply guessed at. Sharpness is not accuracy. Detail is not authenticity. Sometimes leaving something slightly unresolved is actually more faithful to the source than producing a beautifully detailed version of something that was never really there.

I also want to be very clear about what images I use this process on, because I don't want anyone assuming I am doing this to the photographs in my archive. I use it only for Charlie’s Angels images taken directly from the television series on Blu-ray. Those frames are the exception. My other images are authentic photographs and are not being AI-restored, reconstructed, or recreated. I am not taking original photographs of Farrah and running them through AI to improve them. The Blu-ray images begin as frames from the television series, and I want to see how much I can clean and clarify those frames while keeping the original photographed moment intact. This is an incredibly tedious process. 

That also means I am not trying to make Farrah, Jaclyn, Kate, Cheryl, or anyone else look better. I am not interested in improving somebody’s face, fixing an expression, or creating a more glamorous version of what appeared on screen. If the face changes, the process has already failed for me, no matter how technically impressive the rest of the image looks. The same applies to the pose, hair, lighting, composition, and overall atmosphere of the scene. I want the finished image to remain recognizable as that exact moment from the series, not an AI interpretation.

That is also why I don't see this as inconsistent with my pushback against AI images involving Farrah. I have been very critical of images that manufacture scenes that never happened, create photographs she never posed for, or place her into situations that never existed. That is something else entirely. There is a difference between starting with a real photograph and trying to preserve it, and starting with a real person and asking AI to invent a new moment around them. Both may use the same basic technology, but they serve different purposes.

Even with the restrictions I use, I don't automatically assume a result is good just because it looks good. I compare it with the original frame. Sometimes an error jumps out immediately. Other times it is something much smaller. An expression may have shifted slightly. The hair shape may be different. A face may look just a little too polished. A shadow may have changed enough to affect the image's character. Most people might never notice some of those things, but after more than four decades of working with photographs, I tend to notice them. That is probably the most important part of this entire process because the software cannot make that judgment for me.

AI is a tool. It can do extraordinary things, but it doesn't know which parts of an image matter to me unless I control the process and judge the result myself. It can make something cleaner. It can make something sharper. It can certainly make something prettier. None of those things automatically mean it's more faithful to the original.

For me, that is what this comes down to. I am not interested in seeing how much AI can add to an old Charlie’s Angels frame. I am interested in seeing how much I can improve the way that frame is presented while changing as little of the original image as possible. The technology makes invention incredibly easy. The challenge is knowing when not to let it invent.