I’d like to give a very special thank you to Sandra Cannon, I was first introduced to Artificial Intelligence through her beautiful AI artwork last year, seeing new and interesting things emerge from her prompt engineering journey with Mid Journey was a real treat. This was the catalyst for my deep dive into AI language models, and specifically those trained for Art generation. I’m very grateful for this introduction, thank you Sandra. I find that AI is an endless playground where I’m able to unify of all my creative skill sets, and dream up entire worlds.
My process began in Mid Journey. Here I explored long prompt engineering sessions to get the scene concept fine tuned to my direction. Once I had a repeatable prompt and artwork from Mid Journey I turned my focus to Blockade Labs. Here I uploaded my generated artwork to train AI on the image data. Then I began to generate scene variations based on my concept art. Once I generated a scene that achieved my vision I brought the generated equirectangular into an AI upscaler by Topaz Labs. This export was brought into Lightroom where additional details, highlights, curves, color, texture, and sharpness were applied.
With a clean sharp 24,576 x 12, 288 pixel image export from Lightroom, the fun could begin in photoshop. With the background layer under the overlay grid in Photoshop I could begin compositing individual high resolution objects and details for the scene back in Mid Journey. These AI generated objects and images also went through up-scaling, and adjustments in lightroom to match the scene. Once imported as smart objects in Photoshop they would go through a rectilinear to equirectangular projection conversion through smart object warping.
I found the key to this workflow was visualizing objects one might find in a futuristic Cypherpunk scene in the year 2081. Once the idea was there, Mid Journey could be used to create some absolutely incredible detail for my scene. After mastering the color, position, and lighting of the object, it would fit seamlessly during the compositing phase.
Creating a 24,576 by 12,288 equirectangular projection with AI presents a unique set of challenges. While AI technology can be used to generate and upscale images, it requires significant computation and fast drive R/W speeds to work with such an enormous image. Editing in Photoshop and Lightroom at this scale is quite challenging even with the most current hardware configurations.
For this project I used a maxed out M1 Ultra Mac Studio and even this machine was sluggish at times. Additionally, training the AI models to generate and upscale images at this scale requires a large amount of high-quality training data. Ensuring that the AI-generated image accurately represents the original scene and is free from artifacts or distortions was difficult to say the least.
In order to edit in equirectangular space you need to change the image projection when importing a rectilinear image, this conversion needs to be done in a convincing way, so a grid overlay is used. High resolution rectilinear still images would be imported into photoshop converted to a smart object and then warped through Photoshop. For objects with heavy distortion fields near the Zenith and Nadir I placed and re-projected them with Pano2VR.
The Grid Overlay seen here represents distortions of a rectilinear view when projected in equirectangular space. Using the overlay helps keep distortions at a minimum when in VR and when viewing the rectilinear projection of the equirectangular image.
Carefully crafted prompt engineering is essential to the success of any AI generated project. Those who dismiss AI image generation as a simple prompting exercise do not understand the process of directing AI to very specific constraints. Lighting, color, depth of field, style, subject position, camera angles, seed numbers, and negative prompting, are all crucial to create a seamless composite across all images. Countless hours of practice with AI, and the concise economy of words used are required to keep the constraints consistent across all renderings.
Even with specific prompting you will undoubtedly need a healthy dose of patience and some re-rolling of prompts.
Example mid journey prompt: cute cyberpunk kitten curled up a red chair, red futuristic outfit, teal blue eyes, purple hair, back lit, dim lighting, front view –ar 3:2 –q 2
As I mentioned at the beginning, I owe a great deal of thanks to Sandra Cannon who first introduced me to AI language models through her artwork in 2022. I watched her battle cancer with creativity and art. It left a powerful in-print on me that I will never forget. Sadly Sandra and I never got to meet in person, but towards the end of her life she had begun to reach out and connect with me. Tragically Sandra passed away the same day I completed this project. I will never know if she received my message of gratitude so I’ve immortalized her in a tribute version of this image with the last photo she ever posted on facebook. Sandra Cannon will live on in our hearts and memories forever.
My deepest condolences to those who knew her and who miss her greatly,
With <3 -TC
