Reproduce “High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation” 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
Dongyang Liu, Ruoyi Du, David Liu, Dengyang Jiang, Liangchen Li, Qilong Wu, Zhen Li, Steven C.H. Hoi and 2 more
Z-Image Turbo++ advances high-fidelity image generation by distilling an 8-step teacher model into a 2-step student, significantly reducing inference computations. This is achieved through distribution-aligned adversarial learning, step-decoupled parameterization, and end-to-end training with iterative regularization, enabling the model to produce detailed images while outperforming other 2-step baselines.
Key insights
- Using teacher-generated images for adversarial training, rather than real images, significantly stabilizes optimization and reduces artifacts in ultra-few-step distillation by providing a more aligned target distribution.
- Decoupling parameters for each step in a 2-step generator is crucial for effective task specialization and overcoming capacity bottlenecks caused by highly distinct denoising responsibilities.
- Retaining an intermediate (step-1) loss as iterative regularization during end-to-end training is essential for preserving the pre-trained model's progressive denoising knowledge, preventing quality degradation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.