RenAIssance: A Survey Into AI Text-to-Image Generation in the Era of Large Model

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Text-to-image (TTI) generation has evolved significantly, with diffusion models becoming key image decoders.
• Advancements in large models and language model integration have dramatically improved TTI performance.
• Current TTI models produce results nearly indistinguishable from real-world images, transforming image retrieval.

Purpose of the Study:

• To survey and detail major text-to-image generation frameworks.
• To provide a comparative analysis and critique of existing TTI methods.
• To identify potential future research directions and improvements in TTI.

Main Methods:

• Review of literature on text-to-image generation models, including GANs, Transformers, and diffusion models.
• Categorization and analysis of different TTI frameworks.
• Comparative evaluation of method performance and limitations.

Main Results:

• Diffusion models are the dominant architecture for high-fidelity image synthesis in TTI.
• Scaling model size and integrating large language models enhance TTI capabilities.
• The study identifies areas for innovation in model architectures and prediction techniques.

Conclusions:

• Text-to-image generation is a rapidly advancing field with potential for significant productivity gains in AI-generated content (AIGC).
• Future work could extend TTI capabilities to complex tasks like video and 3D generation.
• Further research into novel architectures and enhancement techniques can push the boundaries of TTI performance.