Reproduce “HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified 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, Chengmin Gao, Zijian Gong, Yehao Li, Yingwei Pan, Yi Peng, Zhaofan Qiu and 17 more

HiDream-O1-Image uses one pixel-space Transformer for text-to-image generation, instruction-based editing, and subject personalization. It combines text, reference-image conditions, and noisy target-image patches in a shared token sequence, without a VAE for target images. On UniSubject’s 9–11-subject test, its table Q-O score is 7.65 versus Echo-4o’s 6.73, though prose reports 7.48; model-judged results do not isolate architecture as the cause.

Key insights

  • Directly unifying multimodal inputs (pixels, text, conditions) within an end-to-end pixel-level Transformer architecture significantly improves cross-modality alignment and avoids information bottlenecks.
  • Integrating a reasoning-driven prompt agent enhances the model's ability to interpret and execute complex, abstract user instructions, bridging the gap between user intent and generative output.
  • A natively unified pixel-space diffusion model architecture is scalable, demonstrating strong performance across both efficient 8B parameter and high-fidelity 200B+ parameter versions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.