The Development of AI Images: From Abstract Pixels to Lifelike Pictures

Explore the evolution of AI image generation from VAE and GAN to Stable Diffusion. Learn about the benefits of Stable Diffusion in creating high-quality, controllable AI images and how it's revolutionizing AI headshot technology.
BlinkHeadshot.AI's avatar
Jul 26, 2024
The Development of AI Images: From Abstract Pixels to Lifelike Pictures
Every day, new advancements in AI technology are readily accessible to us. Of all these quick developments, the development of AI-generated photographs is particularly noteworthy. Let's take a look at how AI picture production has developed to this amazing point.

In the Beginning of AI Picture, there were VAE and GAN

The Emergence of VAE (Variational Autoencoder)
Imagine a student attempting to create a portrait in order to comprehend VAE:
  • Encoder: The student observes at someone's face.
  • Encoded data (bottleneck): Maybe the student will remember a person’s face like this way:
    • Eyes: light brown / Eyebrows: thick and black
    • Face shape: round, but somewhat resembled to an egg
  • Decoder: Using encoded data from their observations, the student draws the face.
Image 2. Faces Generated by using VAE: https://github.com/wojciechmo/vae
Image 2. Faces Generated by using VAE: https://github.com/wojciechmo/vae
In summary, VAE is like a student who is drawing a portrait: 1) Observe some features, 2) Remember the features in a simple way, 3) Draw portrait based on what the student memorized. As you can imagine, the student wouldn’t draw the portrait as exactly same as the observed person. This indicates some randomness into the process.
This showed that VAE can generate new image following similar to the given data.
 
The Rise of GAN (Generative Adversarial Networks)
Image 3. Generating Practical Adversarial Network Traffic Flows Using NIDSGAN: https://arxiv.org/pdf/2203.06694
Image 3. Generating Practical Adversarial Network Traffic Flows Using NIDSGAN: https://arxiv.org/pdf/2203.06694
GAN is a deep learning model for creating fake images that look real. The inventor of GAN, Ian Goodfellow, compares it to money counterfeiting.
GAN uses two models:
  • Generator: Produces "counterfeit" pictures
  • Discriminator: Determines if the image from the generator is real or fake
Then, the generator improves its "counterfeits" based on the Discriminator's feedback. Generator and Discriminator continue this process on and on.
Image 4. (AI Picture by using GAN) Figure 2: Some examples of GAN generated faces, including DCGAN, ProGAN, StyleGAN, Style-GAN2, StyleGAN3 (left to right columns): https://www.techscience.com/jihpp/v4n1/48446/html
Image 4. (AI Picture by using GAN) Figure 2: Some examples of GAN generated faces, including DCGAN, ProGAN, StyleGAN, Style-GAN2, StyleGAN3 (left to right columns): https://www.techscience.com/jihpp/v4n1/48446/html

Drawbacks of VAE and GAN

Nevertheless, VAE and GAN have a few drawbacks:
VAE Limitations
  • It tends to generate blurry or fuzzy images.
  • It frequently overlooks fine details.
  • It assumes a Gaussian distribution, which might not match the distributions of actual face images.
GAN Limitations
  • Training is challenging.
  • It usually produces visuals that are not very diverse.
  • Compared to VAEs, it demands higher processing power.

Why People Use Stable Diffusion?

If you've been paying attention to generative AI, you've probably heard about "Stable Diffusion." Let's examine its advantages over VAE and GAN.

What is Stable Diffusion?

Image 5. Popular model for AI Image generator (Stable Diffusion): https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
Image 5. Popular model for AI Image generator (Stable Diffusion): https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
Stable Diffusion is a foundation model released by Stability AI, which trained a Latent Diffusion Model and made it publicly available. Latent, similar to how VAE extracts essential information from pixel-based images, is a technique used to compress high-resolution images (such as 1024x1024 professional headshots) into core image information called "latent space". This compression allows for the rapid generation of high-resolution images. This technique is particularly useful in AI headshot generators, allowing for the quick creation of high-quality professional portraits.
Building on this latent space concept, Stable Diffusion employs a unique approach to image generation. A deep learning model called Stable Diffusion is designed to produce data by progressively denoising random noise in the latent space. This process works in reverse of adding noise, gradually refining the latent representation until it can be decoded into a clear, high-resolution image.
It begins as an old TV with no signal, with a canvas filled with erratic static. Imagine tuning an old TV signal to see a clear screen. The noise is gradually eliminated by the model until a distinct image is produced. In theory, stable diffusion learns to add noise to an image in the opposite direction.

Benefits of Stable Diffusion

  1. 🖼️ Image Quality: Stable Diffusion may provide clear, high-quality photographs, in contrast to VAEs, which frequently yield hazy images. It preserves fine details more effectively than many GAN implementations and VAEs.
  1. 🏋️‍♀️ Training Stability: As its name implies, stable diffusion provides a more stable training process than GANs, which can be challenging to train.
  1. 🎨 Flexibility in Generation: Conditional generation is an area in which Stable Diffusion shines, giving users exact control over the images that are produced through text prompts. Reaching this level of control with VAEs or GANs is frequently more difficult.

The Future of AI Headshot: BlinkHeadshot.ai and Beyond

Thanks to the evolution of AI image-generation technology, we are able to construct our AI headshot generator at BlinkHeadshot.ai. We're committed to ongoing research in order to give our customers better quality and greater satisfaction. Take advantage of the most cutting-edge AI portrait technology available by trying BlinkHeadshot.ai right now.

Conclusion

Embracing the AI Headshot Revolution

The journey from VAEs to Stable Diffusion represents more than just technological progress; it's a democratization of professional imaging. Whether you're a remote worker needing a polished LinkedIn profile or a business looking to update your team page, AI headshot generators like BlinkHeadshot.ai offer an accessible, high-quality solution.
 
Plus, see our previous post "How AI Headshot Generators Work and Why BlinkHeadshot.ai Stands Out" for additional information about diffusion models and techniques for producing AI headshots.
 

References

 
Share article
Subscribe to our newsletter.

BlinkHeadshot.ai