"New 'Hourglass Diffusion Transformer' Sets State-of-the-Art for High-Resolution Image Synthesis, Bringing Efficiency and Scalability to Pixel Count Training"

Hourglass Diffusion Transformers

Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers Stability AI1, LMU Munich2, Birchlabs3, Independent Researchers4Preprint, 2024 *Indicates Equal Contribution Teaser of sample images from our HDiT models trained on FFHQ-1024^2 and ImageNet-256^2. Samples generated directly in RGB pixel space using our HDiT models trained on FFHQ-10242 and ImageNet-2562. Abstract We present the Hourglass Diffusion Transformer (HDiT), an image generative model that exhibits...