metadata
pipeline_tag: unconditional-image-generation
FiTv2: Scalable and Improved Flexible Vision Transformer for Diffusion Model
This repository contains the FiTv2 model described in FiTv2: Scalable and Improved Flexible Vision Transformer for Diffusion Model. FiTv2 is a flexible vision transformer for diffusion models that generates images at unrestricted resolutions and aspect ratios. It builds on FiT with innovations such as Query-Key vector normalization, AdaLN-LoRA, a rectified flow scheduler, and a Logit-Normal sampler.
The official code and sampling scripts are available at https://github.com/whlzy/FiT.