Two more Elphabas

 I got inspired to create a couple more Elphaba renderings using the Canva app & my two favorite diffusion-model-generated versions.

I’ll reiterate that I don’t think generative AI/diffusion models are necessarily a viable way to create actual costume designs for a real-world theater production, in which the costumes will be custom-built based upon the renderings. It’s not that it’s impossible, it’s just that the technology does not allow for the kind of aesthetic control that costume designers seem to want, and I think the shady ethics of training model data will probably be insurmountable for most artists/designers.

In all of this research, I’ve been approaching it as an educator, with a mind to the fact that many of us are teaching at institutions encouraging the use of these technologies (or in some cases requiring their incorporation into pedagogy) and there will be student use no matter how restrictive your syllabus AI policy.

Perhaps an assignment idea could be to create a costume rendering using only digital tools like generative AI and digital image editing/drawing apps, and then create the same one using analog media, like pencils and paint, croquis tracing, etc. 

I don’t know, & I don’t teach these kinds of classes. I do know through the grapevine that students at reputable MFA programs around the US and Canada are already incorporating generative AI into their design processes, whether their professors can see that’s what they’re using or not. 

Anyhow, I’m revisiting Elphaba, recalling that one of my design professors criticized renderings with no background to situate the figure in space. Even a horizon line or a vague shadow was preferred.

.

Incidentally, I asked Microsoft Copilot how much energy it takes to generate one image. Here’s what it said:

Generating a single image with Copilot typically consumes about 0.0029 kWh (10.44 kJ) of energy 1. This is roughly equivalent to charging a smartphone battery to 24% 1. The energy consumption can vary depending on the complexity of the image and the model used.  

1. Futurism.com 

I asked it what factors influence that variability.  

Several factors influence the energy consumption of generating images with AI:

• Model Complexity: More complex models with a higher number of parameters require more computational power, leading to increased energy use 1. 

• Hardware Efficiency: The type and efficiency of the hardware used (e.g., GPUs, TPUs) can significantly impact energy consumption 2. 

• Image Resolution and Detail: Higher resolution and more detailed images require more processing power, thus consuming more energy 1. 

• Optimization Techniques: The use of optimized algorithms and techniques can reduce the energy required for image generation 2.  

• Cooling Systems: The energy needed to cool the hardware also contributes to the overall energy consumption 2.

Understanding these factors can help in making more energy-efficient choices when using AI for image generation.

1: MIT Technology Review 2: MIT News

Someone want to create a browser plug-in with that data?