CS CASE STUDY KEY TERMS AND DEFINITIONS

Bias Mitigation

Techniques and tools used to reduce unfair or inaccurate patterns in an AI model caused by biased training data.

Character Consistency

The ability of an AI image model to generate the same character with similar appearance across multiple images.

Convolutional Neural Network (CNN)

A type of network model designed to process images by detecting features such as edges, shapes, textures, and objects.

Dataset Curation

The process of selecting, cleaning, and organizing data for use in machine learning models. Proper dataset curation ensures high-quality, relevant data that improves model performance and reduces bias.

Denoising

The process of removing noise from a dataset or signal to enhance the quality of data, often used in image generation. In machine learning, denoising can improve model accuracy and generalization by focusing on the true underlying patterns.

Denoising Diffusion Probabilistic Model (DDPM)

A diffusion-based AI model that learns to generate images by gradually removing noise from a random noisy image until a clear image is formed.

Diffusion Model

A generative AI model that creates new images by starting with random noise and progressively transforming it into meaningful content.

Embedding-Based Approach

A method that converts information such as words or images into numerical vectors so that AI models can understand relationships and similarities between them.

Image Generation

The process of using AI to create entirely new images based on learned patterns from training data.

Conditional Image Generation

Image generation where the AI is guided by additional input, such as a text prompt, sketch, or reference image.

Unconditional Image Generation

Image generation without any user-provided guidance; the AI creates images solely based on patterns learned during training.

Image-to-Image Translation

A technique where AI transforms one image into another while preserving key information, such as converting a sketch into a realistic photo.

Noise Injection

The deliberate addition of random noise to data during training or generation to improve model robustness or support diffusion processes.

Segmentation Map

An image in which each pixel is labeled according to the object or region it belongs to, helping AI understand image structure.

Text-to-Image Generation

The process of generating images from written descriptions using AI models that connect language and visual concepts.

Training Stability

The ability of an AI model to learn consistently during training without producing unpredictable results, errors, or failing to