VFX Atlas
H 299 / 3212010s

GAN Portraits

2014, Ian Goodfellow's Generative Adversarial Networks

Typeface · Jura

Synthetic Smiling Man. A smiling young man appears in a tightly cropped portrait against a softly blurred outdoor setting.
Synthetic Smiling Man. A smiling young man appears in a tightly cropped portrait against a softly blurred outdoor setting.StyleGAN2 · Public domain · via Wikimedia CommonsView source ↗

Prompt

Apply a GAN Portraits effect with polished photorealistic facial rendering, soft studio-like illumination, finely synthesized skin texture, and unusually smooth tonal transitions. Introduce subtle generative imperfections in peripheral details, such as uneven accessories, ambiguous hair strands, or gently melted background textures, without changing identity. Keep the original subject and composition.

Paste it into LetzAI with your own image to apply the effect ↗

Identify

You see highly realistic, photographic portraits of human faces. Look closely to spot subtle distortions like melted backgrounds or asymmetrical earrings.

Principle

Two neural networks1 compete to generate and evaluate images. They repeat this until the artificial image passes for a real photograph. The system learns visual rules from real photos but lacks actual 3D understanding.

Example

Tero Karras et al. (NVIDIA): A Style-Based Generator Architecture for Generative Adversarial Networks2 (StyleGAN), 2018.

Try it out

Visit the website 'This Person Does Not Exist'. Look at the edges of hair or glasses for blurry glitches.

Background

In 2014, researcher Ian Goodfellow invented Generative Adversarial Networks (GANs). A GAN uses two computer programs playing a game. The generator tries to forge an image, while the discriminator3 tries to spot the fake. They train together, improving rapidly. By 2018, NVIDIA engineers released StyleGAN. This system could produce incredibly high-resolution, photorealistic human faces. It analyzed thousands of celebrity photos to learn the statistical arrangement of human features. It did not copy any specific person. Instead, it mathematically combined learned features to invent entirely new, non-existent people.

Distinction

Traditional 3D CGI models are built manually by artists using actual spatial geometry. In contrast, GAN portraits are generated entirely by statistical algorithms mapping 2D pixels.

Aftermath

It teaches us that computers can convincingly fake reality by mapping mathematical patterns rather than understanding physical space.

Further examples · 7 images

Synthetic Woman With Brown Hair
02 Synthetic Woman With Brown HairA smiling woman with brown hair fills the frame against a dark background. Convincing skin and facial features coexist with tangled hair boundaries and unresolved forms at the shoulders, revealing the characteristic uneven coherence of a GAN portrait.StyleGAN2 · Public domain · Wikimedia Commons
Synthetic Man With Distorted Ear Accessories
03 Synthetic Man With Distorted Ear AccessoriesA short-haired man with facial stubble appears against a blue-and-black background, with thin cordlike forms beside his ears. These uneven, unresolved accessories contrast with the convincing face and visibly demonstrate the local inconsistencies typical of GAN portraits.Owlsmcgee · Public domain · Wikimedia Commons
Synthetic Woman With Straight Bangs
04 Synthetic Woman With Straight BangsA young woman with straight bangs faces the viewer before a gray, stone-like backdrop. Realistic eyes and skin contrast with merging hair, jacket edges, and background details, showing the persuasive surface realism and spatial inconsistencies of GAN portraits.Owlsmcgee · Public domain · Wikimedia Commons
Synthetic Boy in a Collared Shirt
05 Synthetic Boy in a Collared ShirtA boy wearing a pale collared shirt faces forward against a dark background. The photographic likeness and subtly irregular teeth and ear contours provide a restrained example of the GAN portrait effect.Owlsmcgee · Public domain · Wikimedia Commons
Synthetic Portrait Against Brick
06 Synthetic Portrait Against BrickA young person with long reddish-brown hair appears before a brick wall. The realistic central face contrasts with less coherent hair, clothing, and peripheral masonry, consistent with GAN-generated portrait imagery.bod lnga klang · Public domain · Wikimedia Commons
AI illustration: Synthetic Human Face Generated by a GAN
07 AI illustration: Synthetic Human Face Generated by a GANThis image presents a strikingly realistic human face with hyper-detailed skin textures and lifelike eyes. Upon closer inspection, the mismatched earrings and slightly amorphous background details reveal its origin as a Generative Adversarial Network output.Illustration generated for the Atlas (AI)

Question

Why do GAN portraits often have mismatched earrings or strange backgrounds?

Show possible answer +

The AI only learns 2D pixel patterns, not 3D object rules. It simply does not know that earrings come in matching pairs.

Sources