August

 

                     Like it or not, the way things were is not how things will be.

First, a note for any film industry people in LA who are suffering from the lack of production these days: The LA Times recently published an article (free to everyone) detailing how and where to get financial help. If that's you, please read it and avail yourself of whatever aid that you can. Assuming the Teamsters okay the recent Basic Crafts Agreement, the threat of a labor stoppage will be over, and hopefully Hollywood will get back to work again, but even in the best of circumstances, that won't happen overnight ... and we don't know if or when the town will truly get busy.  Meanwhile, keep the faith -- I wish the best of luck to you all.

Now, on with the show...

Those of you who've been here for a while will remember Peter McLennan, a retired DP/Director who's written three excellent (and very popular) guest posts for this space in the last few years -- but if you missed them, you can rectify that oversight here and here.  He's a terrific writer who's probably forgotten more about the art of putting images on film than I ever knew, which is why I'm always grateful for his contributions to this blog.  A gifted photographer with a great eye developed during his long career looking through the viewfinder, Peter recently began experimenting with images manipulated by an AI program called Stable Diffusion.  He's been sending me the results for a while, so I finally asked "Why don't you write another guest post on the subject?"

It was a win-win proposition: not only would it be a good read for all of us, but it would take the heat off me for another month ... and so it has.

Peter, the floor is yours:  

         A Camera for my Imagination

A retired DP finds new avenues of creativity with Generative AI

Story and images by Peter McLennan

On a dark winter day in 2023 I typed the word “tree” into a text box and watched gobsmacked as an image of a tree appeared on my screen. It was a moment as memorable as when I first watched an image appear like magic in a darkroom developer tray.

My very first AI Image wasn’t much to look at.  In fact, looking back at it eighteen months later, it looks positively primitive.  But for me on that winter day, it was transformative. The image was generated not by a camera or a paintbrush or even by an artist. It was generated by a single word.

I soon discovered that a forest was as easy to generate as a tree, and an artistic rendering of a forest was as easy as adding the words “etching” and “art deco” to the prompt. As a constant stream of images appeared on my screen, a trap door opened on the floor of my office and I fell through it, not emerging until many weeks later. Day after day I spent hours exploring, working, making “just one more generation” until one day I realized what had happened. I had become addicted to Generative AI.

In a mad burst of creativity and discovery I began to explore a new world of images both familiar and novel. I made images that were horrifying and beautiful, intentional and accidental, disappointing and delightful and always, images that were surprising.  

I think it was the discovery aspect that kept me glued to the keyboard late into those winter nights, for I frequently had no idea what would appear on my screen. Many of the software’s controls achieved their results with completely new concepts using new and arcane names like “Variational Autoencoder”, “Denoising Strength” and “Classifier Free Guidance”. I had little idea what any of them meant, so the only way forward was to experiment. Consequently, many images were far from what I expected, and as it turns out, surprise and delight were what kept me engaged with Stable Diffusion.

My initial discovery process used what is called “text to image” where the images are generated solely by your skill at writing prompts.  For me, as a life-long photographer, this was challenging.  I was used to seeing images and recording what I saw in a viewfinder.  Now, I had no viewfinder. I had only my thoughts.

I tried without success to illustrate an ocean voyage I’d taken on a freighter as a teenager, but the images on my screen bore no resemblance to those in my memory. I failed completely to illustrate what I’d intended.  I felt I’d plumbed the depths of generative AI and found severe limitations.  Had I reached the end of my creative buzz?

In fact, I’d barely begun.

Soon, another technique presented itself: Image to Image.  Now, I could take existing images, my images, and operate on them using the same AI techniques as before.  Another new world opened up. For instance, this rather boring photograph of a door in Nevada …

… became this one.  Bearing only vaguely visible resemblance to the original, the camera image has become something completely new.

A photo I took in Sri Lanka…

... became this, from dog knows where. Who are those guys?

This image -- part photograph, part creative prompting, and part styles AI learned from artists who worked long ago -- started life as a simple photo of snowy trees in my backyard.

The “image to image” function effectively divides the image creation process into two components: The AI, and me. My photography provides the basic structure and palette, Stable Diffusion interprets my text prompts and adds its own creative seasonings as it sees fit.  

It is a productive and inspiring, if somewhat hit and miss, partnership. This image of confidence, youth, fearlessness and curiosity was generated from a photograph of some construction machinery in a grassy field.

The AI software doesn't need to work from a photo, drawing, or painting. These next four images were created entirely from my imagination using prompts to describe what I wanted to see, including the media type, image size in pixels, and aspect ratio. Most of these prompts resemble descriptions found in film scripts, but some trial and error was required to jump the gulf from imagination to screen image. The resultis were further controlled by several on-screen math functions related uniqely to the Stable Diffusion imaging software. 

“Oil painting, messy Victorian library, winter afternoon, dramatic window lighting, a body is on the floor, a man sits at a distant desk, volumetric lighting”

“High resolution colour photograph, night, interior truck cab, rain”

“Cartoon colour image of an orange cat flying a kite in a field of flowers”

“High resolution monocrhrome photo of a film set from the forties, smoke, movie lights, several crew members”

The astounding pace of AI development in the last two years points to an uncertain future. Although we can't predict exactly when and how it will impact the film and television industry, it’s already apparent that the creative skills in immediate peril are those of photographers and illustrators. But how will musicians, film crews, and writers fare in the new world of AI?  Repetitive clerical and administrative tasks are far more efficiently done by AI, and the writing is on the wall for copywriters and computer coders. If I was an office worker tasked with researching and writing reports, I’d be brushing up my resume -- presumably with the help of a bot of some kind -- and I’d be learning everything I could about the coming tsunami of AI assistance.

That’s the most important thing I’ve learned with my admittedly shallow-end dip into the pool of AI imaging: AI doesn’t do the work, it helps you do the work. AI doesn’t think for you, it assists you by doing much of the grunt work, freeing you to be more creative, more explorative, more adventurous. A far better term than “Artificial Intelligence” would be “Assistive Intelligence." I didn’t need to learn how to paint with egg tempera or do etching or lithography, my AI pal knew all that. I just had to tell it what I wanted to see and what artist’s technique to use. 

More impressive, AI seems to think for itself, drawing on resources I’m ignorant of and regularly coming up with ideas I’d never have considered.  The elusive muse we call creativity works in mysterious ways. As this technology becomes more adept and accessible, I foresee a bright future -- not for computer coders, storyboard artists. and report writers, but for audiences.  We’re in for a treat.  

Just you wait.

PS: This experimental clip -- not affiliated with Volvo in any way -- is a sample of what AI can do now ... so imagine what it will be able to do in five years.