Tip Of AI
Discover what you can do with AI
HiDream-E1 is an open source instruction based image editing model built on top of HiDream-I1 foundation offering high efficiency and quality image editing. It uses a sparse diffusion transformer and supports instruction driven editing with dynamic resolution up to one million pixels. It features a refine_strength parameter that allows users to balance between the main editing operation and final refinement in the pipeline. It requires Flash Attention, CUDA 12.4, and the Diffusers library for efficient execution. The model provides two variants: HiDream-E1-Full with fixed 768x768 resolution and HiDream-E1.1 with dynamic resolution support and improved image quality. Both variants use 28 inference steps and can automatically download required model files from HuggingFace when the user has authenticated their account. Supported workflows include direct instruction input for dynamic editing, image to image refinement, interactive Gradio demos, and command line scripts. Evaluation on EmuEdit and ReasonEdit benchmarks shows high accuracy in color, style, text, and local edits. The MIT license ensures free usage and modification by developers and researchers.