Every living thing has a face. We can have a long list of it from dogs to apes, to reptiles and amphibians, and even insects have it. Nature is so constant in creating the face. It has one formula, a forehead, a pair of eyes and ears, a nose, a mouth, and a chin. Essentially, it maximizes survival. We have eyes elevated to have a better perspective of the environment. Our nose is open downward so that rain can't pour into it and drown us. Our mouth is at an optimal distance for us to intake food.
Every time we meet a person, we see the faces first. Faces can reveal more about the person than any of our body parts. We can have petite and short, fat and bouncy, or an edgy and long face. I think that is what makes us unique. Reconstructing and creating human faces with AI has been on the radar for some AI researchers. It even produces hyperrealistic output faces that we can hardly differentiate from real people. While I was tweaking AI Art, I experimented with creating AI portraits using VQGAN+CLIP, but not easy to achieve what I envisioned until Nsheperd merges CLIP with StyleGAN.
 |
 |
When I was experimenting in VQGAN+CLIP, most of the portraits haven't blessed with a well-defined face. I did countless of experimentation, and a good face feature seems to find a needle in the haystack. I did recommended technique from the write-up of 
Although StyleGan is seemingly better with AI portraits, often it doesn't meet what I envisioned. I started with the target as prompts and a seed of -1. I find the results too close to one another, and I find it boring and repetitive. Sometimes, the portraits just lay away from the prompt. Hence I wasn't satisfied with the AI portrait; I deliberately encoded the prompt as detailed as possible. I explicitly put in my prompts the elements to what I want to see in the AI portraits. I applied weights to each keyword and provided a target image, which I created using Arbreeder, shown below.
 |  |
 |  |
I used four target images in my experiment run, which serve as the genes giving my AI portrait's facial features. The idea behind using multiple target images is to mix and match features like parent genes. I got the idea from tweaking on Artbreeder. To generate a substantially good AI portrait, I performed trial errors in giving weighs to different prompts, making my AI portrait close to what I envisioned.
For the topmost featured illustration in the blog, I did a prompt consisting of keyframes, background, keywords like goddess and woman, and the target images. I used the Keyframes to ensure my portrait lies in the center. To my surprise, it did work well and was consistent in my successive runs. I also tried close-up, and it works well. I was mesmerized by the AI portrait when I did animate it. It seems it is as if it was having metamorphosis while walking like a true goddess.
METAMORPHOSIS
I was really into the concept of metamorphosis for most of the week. Maybe, it is because of the documentary I watched recently about metamorphosis. When we thrive for change to become better all our lives, we forget that the metamorphosis was to return to ourselves. We need to recognize that change is slow rather than what others are telling us to do. The idea rubs off with the theme of the AI portraits, which I was creating.
 |
 |
In the featured artwork above and below, I tried to mimic the concept of metamorphosis. I envisioned having my animation coming from a goddess that transformed into a witch. What I did was retain the same prompt from the first illustration. I reset the weight of each prompt but preserved the weight for the target images or genes. I added new keywords like a witch, queen, and tribe. Also, I included Metamorphosis I as one of the target images.
The result was excellent. It did have the witch vibe, and gold highlights represent a queen. While the preceding illustration, it shows a tribal woman. I added at what tribe she aligned. I did put Amazonian as one of the keywords prompts. I also experiment with keyframes. Instead of center, I placed to Keyframes and once for closed up. When I did the animation, I was particularly in awe of the result. It is as if the AI portrait zooms out and in.
 |
 |
FACES OF WAR
War is the most disastrous of all human endeavors. The experience alters us in many ways. It is not just the civilians caught up in it, but the soldiers who fought. I think war and metamorphosis are closely related. Both talk about change, which the former resolves with violence and disruptive means. War got elicited a wide range of emotions in people. I was looking into an online gallery of the faces of war. The before and after photos were enlightening to watch.
 |
 |
As an extension theme to my AI portraits, I wanted to portray the faces of war. Different wars happened in our history, from nomads war with wilds to war with tribes and war of states. I did a new set of prompts referring to keywords that can describe people who participated in the war. I focused the runs on generating male portraits.
In this experiment, I specified the tribe to which the warrior came. I did settle with Aztec for the warrior, and the tribe chieftain is of the Native-American. I was astonished by the portraits. The bandana on the chieftain is of the native-Americans, while facial tattoos of the warrior are of the Aztecs. These minute details in the AI portraits make the output illustration relevant to that of the prompt.
 |
 |
The next run focused on the generating portrait of modern war heroes. I run a prompt specifying to generate a war general and veteran. The result was good, but I think some inconsistencies with the centering happen, despite the keyframe being similar to the other runs. The War General turns out to be particularly good, with the animation showing how it morphs from the initial image. On the other hand, the War Veteran seems offset in orientation, but I like how it turns out. If we pay close attention to the War Veteran, I find it, at a glance, that the lines seem off, but it resembles the bar on typical police and army uniforms.
 |
 |
 |
AI is our creative collaborator. It enables us to imagine the world from a different creative perspective. As our interest in AI art grows, we are also interested in producing portraits of humans we didn't even know. I find my little experiment fruitful, and I can say AI can paint faces with guidance. We can no longer say AI can't do arts as much as portraits because it does. From a prompt and a target image, we can create faces that we have never seen before. Some were rhetorical, and some were hyperrealistic, which blurs the line between AI and real people's portraits.
| Author's Note: |
|---|
All content including digital illustrations is from the author. The illustrations are generated by Enhanced VQGAN-CLIP, MSE Regulized, Artbreeder and Nsphered's StyleGAN3. For enhancing the resolution, color, and tone, the author used Real-ESGAN, Lightroom, and Photoshop. For animation or video clip, I used Sony Vegas. |