Reproduce “HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang and 14 more
HiDream-I1 from HiDream.ai introduces a sparse Diffusion Transformer architecture that generates high-quality images with significantly reduced computational requirements. The model achieves state-of-the-art prompt adherence and human aesthetic preference while producing images within seconds.
Key insights
- A hybrid approach combining dual-stream and single-stream Diffusion Transformer (DiT) blocks with dynamic Mixture-of-Experts (MoE) modules allows for scalable capacity and efficient processing of image and text modalities.
- Multi-stage training, including progressive resolution and fine-tuning with high-quality human-annotated data, optimizes both fundamental diffusion process learning and human aesthetic preference alignment.
- Knowledge distillation, including GAN-powered distillation, effectively creates faster model variants (HiDream-I1-Dev, HiDream-I1-Fast) that maintain perceptual quality while significantly reducing inference time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.