/imagine

July 1, 2024 artificial intelligence

I’ve always wanted to write blog posts, mostly to document my journey through ai. However, with work (I'm a freelance photographer, 3D artist and recently AI artist), personal projects and 2 young kids taking most of the time available in any waking day (i’m trying to write this with my 1.5 year old running and screaming in circles around the dining room table), that project was evidently put on the back-burner for the better part of the last 18 months.


So here I am, finally drafting this first post, trying to gather up whatever operating neurones i have left today (it's noon, but feels like 11PM) to explain how I came about creating a piece i’ve started long ago and only recently really called "done". Not only is it an expression of things i cannot otherwise say out loud (something something about art and soul stuff), I regard it as an interesting personal checkpoint as it combines skills and knowledge i've gathered at both ends of the journey so far, the beginning and today. 


When I made it initially, I was only beginning to scratch the surface of Midjourney (that I would later drop) and Stable Diffusion (which will become my defacto creation ground). I knew almost nothing of what MJ prompting was about, or the breadth of tools to be available to me with SD. Fast forward to today and I'm eye balls deep in SD tech, from custom nodal networks in ComfyUI to training LORAs in Kohya SS and inpainting 100 megapixel images in Fooocus. All done locally on a system I built and manage myself. So it's fun, perhaps to me alone, to see how things were to where they are at today. And this pieces sort of reflects that very vividly.

Early generations (Midjourney v4)

Why Schism


It is borrowed from a track of the same title, written and performed by metal band extravaganza Tool. I used to hammer the repeat button on their CDs throughout my twenties and thirties (I'm turning 42 this November), with the Lateralus album probably ranking the highest on my personal charts.


To those who have never heard it, it’s a song about two lovers experiencing the devastating effects of terrible communication (and the lack of it), ultimately breaking apart in an irreparable way. Each time I listened to it, I kept visualizing towering monuments of humans splitting along the seam that previously united them. Or crashing into each other. Or simply eroding away. You get the idea.

Overview


  • Part I (January 2023): It all started with a basic Midjourney generation which was then interpolated with a 3D scene made in SetALight (I use it for previzzing client photoshoots, quite nifty and leaner than say Blender or Unreal). That is then followed by loads of different interpolations (img&text as well as img&img) in Midjourney to find a proper base. It took around 30 rolls and 140 generations to get to it. Said base image was then sent to Invoke AI (running a Stable Diffusion 1.5 model), where it underwent over a thousand different inpainting generations to get to a final render I thought was satisfactory at the time.


  • Part II (June 2024): The image is still one of my favorites of all time, I love the concept but the execution leaves much to be desired not even 1.5 years later. it's not satisfactory at all any more. It lacked definition and any kind of proper detailing compared to what I could output these days. I also had just stumbled on Fooocus, which was highly lauded for its inpainting capabilities. I thus wanted to put it to the test, and Schism was the excuse. First, some upscaling prep in ComfyUI, my StableD software of choise. Thanks to this, I was able to properly scale and cleanup the image with sufficiently clear details now available so that I could work on them through more inpainting in Fooocus (this time using an SDXL model). 1040 generations, half a dozen upscales and countless back and forths to Photoshop later, a monumental 12K render was achieved. A quick send to Capture One for a last grade and it was called done.


Following me still? Awesome. Let's dive into the details. Part I is up next.