How are AI models trained, and what are the biases?

Question

Accepted Answer

AI Models Are Trained on Image Banks That Can Reflect Biases

Generative artificial intelligence models used in fashion are trained on extensive banks of images. These datasets often contain retouched photographs and can reflect a dominant "white, Western" aesthetic, potentially leading to biased outputs. While AI can generate a characteristic face type that varies between different platforms, such as MidJourney's tendency to produce models with thicker lips, the underlying training data significantly influences the final imagery.
Training Data Composition

The foundational training for generative AI in image creation relies on large collections of photographs. These image banks are frequently curated with retouched visuals, which may not represent a diverse range of appearances. Furthermore, the existing datasets often exhibit a bias towards a specific look, predominantly aligning with a "white, Western" standard of beauty.
Emergence of Characteristic Aesthetics

A notable characteristic of AI image generation is the tendency for each specific generator to produce a discernible facial type. For instance, the MidJourney AI platform has been observed to frequently render models with fuller lips. This phenomenon suggests that the algorithms, while capable of variation, develop a signature style influenced by their training data.
Addressing Generic Results and Biases

To counter the creation of overly generic results and to mitigate inherent biases, practitioners often refine the descriptive prompts provided to AI engines. Some professionals also actively work to diversify the training data by incorporating "imperfect" databases. This can involve adding images that feature more common skin tones, such as those of individuals known to the trainers, aiming for a more authentic representation.

Reading history

Hide →

This history is cleared when the page is closed.

How are AI models trained, and what are the biases?

Loading

Was this article helpful?

Helpful

Not helpful

Report an issue

2

0

0

Share with friends