Dart: Denoising Autoregressive Transformer (arxiv.org)

A new model called DART (Denoising Autoregressive Transformer) has been proposed to enhance text-to-image generation, addressing the limitations of traditional diffusion models that rely on a Markovian process. DART integrates autoregressive techniques with a non-Markovian framework, enabling it to iteratively denoise image patches both spatially and spectrally. Notably, DART forgoes image quantization, which enhances its modeling capabilities and flexibility while allowing it to train on both text and image data within a unified framework. This advancement is significant for the AI/ML community as it introduces a more efficient and scalable approach to image synthesis, setting new benchmarks for performance in class-conditioned and text-to-image tasks. By leveraging the same architecture as standard language models, DART promises to deliver high-quality images while reducing inefficiencies typically seen in training and inference. The implications of DART could reshape workflows for developers and researchers in the field, providing a robust alternative to existing models and potentially accelerating the development of AI-driven visual content generation.