elevAIte
↝ Description
While I don't think that CG is going to be replaced by AI completely anytime soon, it is an incredibly powerful tool to have in your arsenal; not replacing 3D renders, but aiding and enhancing the process.
Creating intricate backgrounds for 3D scenes tend to be a laborious process, especially when given further restrictions by the client. This is where I helped out: I crafted two workflows, one for generating a background that fits the scene and its style, and another to 'enhance' the details of the final render - especially noticeable in the foliage and the water, both of which can be quite tricky to get right in pure 3D. This allowed the client to choose from a palette of backgrounds and to implement revisions quickly and efficiently.
Another thing to note is that this process was done with print-ready resolution in mind: That is to say, 8k x 8k pixels.
↝ Tools used
ComfyUI & other AI tools
Enhance!
Create!
Another idea I pursued for this project was whether it is possible to delegate quick concept designs of a given - branded - mascot to AI. I was handed only 8 images of the character in various costumes, and I did my best to fine-train a model with those to be able to generate new ones. The results were all we could have asked for: The character was recognizable, the 'costume' you could prompt for could be nearly anything - again, there were only 8 original images, whose costumes I did not choose to generate again here - and the style fit the branding perfectly.
Another thing to note is bias: you might notice the character with the pink, liquid cape; AI tends to try to make liquid blue, especially when the training data is of a blue character. However, the model performed exceptionally well in this regard, and liquid capes could be generated in any color.
Now, AI image generation is a little beast to handle. And while it has gotten a ton easier since the days of Disco Díffusion in early 2022, it still struggles to place objects in the scene correctly based solely on the prompt. What I did below was to create a workflow in which you could sketch out in different colors where the object should be in the image; which of course isn't perfect by any means, but it helps tremendously in the layout phase - especially when you're working with a client who has a very specific vision in mind.
If you add a depth pass from a quick blockout in 3D, you can get pretty consistent and yet varied results for the kind of layout you might have in mind.