A Survey of Text-to-Image Diffusion Models in Generative AI
Siddharth Kandwal, Vibha Nehra
Abstract
From dreamscapes to photorealistic portraits, text-to-image generation pushes the boundaries of AI creativity. This survey navigates diverse techniques, such as GANs, VAEs, and Diffusion models, uncovering their potential for transforming textual descriptions into captivating visuals. These models have significantly advanced the field, but they are not without their limitations. One notable issue is data bias, which could potentially lead to a deficiency in variety and cultural awareness in the produced visuals. Furthermore, recognising the significance of mitigating data bias in generative models, this report offers insights and strategies to address this pressing issue. It explores approaches that leverage inclusive datasets, fairness-aware training techniques, and ethical considerations. These methods aim to bridge the gap between the technological advancements in image generation and the imperative need for inclusivity and cultural sensitivity.