HiDiffusion: Dramatically Faster Image Generation for Diffusion Models 131

HiDiffusion dramatically enhances the image generation quality of diffusion models and boosts generation speed by 6 times. Developed by the research team at ByteDance, HiDiffusionrequires no training and can improve the clarity and generation speed of SD-generated images with just one line of code. It supports increasing the image generation resolution up to 4096×4096, while boosting image generation speed by 1.5 to 6 times. It can also reduce inference time by 40%-60% in high-resolution image generation, and supports all SD models as well as downstream models of SD models, such as Playground.

HiDiffusion Features:

• Supported Models: HiDiffusion supports multiple Stable Diffusion models, including Stable Diffusion XL, Stable Diffusion XL Turbo, Stable Diffusion v2, and Stable Diffusion v1. In addition, it also supports downstream diffusion models based on these models, such as Ghibli-Diffusion and Playground.

• Supported Tasks: HiDiffusion supports multiple tasks, including text-to-image generation, ControlNet control network tasks, and Inpainting image restoration.

• Ease of Use: Through the Huggingface Diffusers library, using HiDiffusion to generate images is extremely simple—only one line of code needs to be added.

• Through extensive experimental verification, HiDiffusion not only solves the problems of object duplication and high computational load, but also achieves unprecedented results in the task of generating high-resolution images.

HiDiffusion Test Address:https://huggingface.co/spaces/multimodalart/HiDiffusion