Photographic vs Photorealistic in the context of AI
I’m writing this post to illustrate the difference between the terms “photographic” and “photorealistic” in the context of AI generative art. I will be using example images from Stable Diffusion checkpoint models that I use in my works.
I’ve prepared the following video so you can see these examples on top of each other, so it’s easier to understand.
Photographic
In AI photography, the goal is to create an image that looks like a photograph — that is, one that was taken by a camera. Images produced this way will largely adhere to the physical properties of light, and depict the subject in a realistic manner that’s possible by the laws of physics.
Virile Reality
Virile Reality, created by Scratchproof, is one such example.
Photorealistic
In the context of AI generation or synthography, photorealistic is a term that has the shading and texture similar to photography, but the depiction of light and optics is not necessarily realistic or reflective of the physical world.
Many of my works fall into this category as I find photorealistic models that are near but not quite photographic to give the most flexibility in artistic interpretation and expression. Here are some examples that you’ll find in my works.
I have arranged these photorealistic models in an order of how close and far they are to being photographic. Sometimes it’s hard to truly show this progressively since a lot of this is quite subjective — but as a rule of thumb, look for how the colors are produced, how the light and shadows interact, and how much texture you could find in the skin — and they smooth or rough, etc.
Airfuck’s Brute Mix
Airfuck’s Brute Mix, created by Airfuck.
Airfuck’s Wild Mix
Airfuck’s Wild Mix, created by Airfuck.
Virile Fusion
Virile Fusion, created by Scratchproof.
Virile Fantasy
Virile Fantasy, created by Scratchproof.
Virile Animation
Virile Animation, created by Scratchproof.
Virile Motion
Virile Motion, created by Scratchproof.
Technical Parameters
It’s really important for me to emphasize that although I was able to generate these images in a very high degree of similarity to one another, I have forced them to be this way by employing a control net using the render made with Virile Reality. In a normal flow, you would never first render with one model and then use that as a control net just to recreate them in another look and feel. However, I did it for these images because I wanted to show the difference between the two terms, and give examples of how the checkpoint models differ in terms of stylistic qualities, so I have forced their forms by using control nets.
Text prompts in Stable Diffusion txt2img with control nets.
- 20 steps, Euler a, 512x512
- Hires 2x, 10 steps, 8x_NMKD-Superscale_150000_G
- 0.5 Denoising
- Automatic1111 1.6.0
- Post Upscale: Topaz Gigapixel HQ 4x, 4096x4096
- CN 0: OpenPose Full, Weight 1, Start 0, End 1, Balanced
- CN 1: Reference Only, Weight 0.5, Start 0, End 1, Balanced
- CN 2: Canny, Weight 1, Start 0, End 1, Balanced