Virtual Avatar by AI

I’m a seasoned online content creator with 15+ years of experience making videos and writing blogs across YouTube and other social media. A trend that emerged during the height of the pandemic was Virtual YouTubers, or “V-Tubers.” Having a model made professionally can be costly — so I set out to see if a V-Tuber avatar could be built using only AI.

Inspiration

The inspiration came from a cyberpunk-styled cosplay helmet purchased from Amazon a year prior. I wanted the character to have a cyberpunk theme with demonic-inspired elements.

Cosplay helmet inspiration
AI Tool

This process uses DreamStudio by Stability.AI, powered by the SDXL 1.0 model. The tool supports multiple visual styles; for this project I chose the Anime style.

AI generated virtual avatar
First Iteration

Using the helmet as a sample image, I fed it into the model with the prompt: FRONT VIEW OF A MAN WITH ROBOTIC ARMS WEARING A FUTURISTIC HELMET WITH GLOWING RED HORNS. SLEEVELESS DARK RED TRENCH COAT AND A BULLETPROOF VEST. FOUR IRON CHAINS HANG FROM HIS NECK LIKE JEWELRY. The result proved too sample-dependent, so I removed the image and relied on the prompt alone.

First results 1 First results 2 First results 3 First results 4 First results 5 First results 6 First results 7 First results 8 First results 9 First results 10

This result turned out poorly. Originally, I thought to just build the rest of the model around the sample-inspired images, but that proved to be difficult since the AI clung too closely to the sample. So I got rid of it and relied on the prompt to build the avatar from scratch.

For a V-tuber avatar, I needed a half-body shot to rig up to the software that gave it its 3D-like appearance. Using the first prompt. I ran another iteration. The results were closer to what I wanted, but a few tweaks to the character’s ethnicity needed to be made.

Second results 1 Second results 2 Second results 3 Second results 4 Second results 5 Second results 6 Second results 7 Second results 8 Second results 9 Second results 10
Second Iteration

The prompt was altered to suit the needs and design of the model I wanted: front view of a dark-skinned man with robotic arms from the waist up wearing a futuristic helmet covering his face with glowing red horns on his head. He's wearing a sleeveless, dark red trench coat and a bulletproof vest underneath. Four iron chains hang around his neck like jewelry.

Second iteration 1 Second iteration 2 Second iteration 3 Second iteration 4 Second iteration 5
Final Result

These latest results were more in line with what I wanted. The prompt was good enough to keep running through it and randomizing the results with new iterations until I picked out a final set.

Final result 1 Final result 2 Final result 3 Final result 4
Post-Mortem

After selecting the final generated image, I hired someone to rig the asset as a live avatar. After debuting the model on YouTube, feedback from the VTubers community showed they couldn’t tell the avatar was AI-generated.Coincidentally, that same person also commented on my use of AI images for the video thumbnails. Funny enough, it seems that they couldn’t tell that the avatar model I used was made with AI as well. There’s probably more I could do to push the prompt to change certain things, like the excess chains around the avatar. But as of now, I think this project and its goal were executed just fine to acceptable results.

Post-mortem avatar
Post-mortem 1 Post-mortem 2