Reproduce “Knowledge Diversion: A Novel Paradigm for Pre-training Decomposable Image Generation Models” 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
Yucheng Xie, Fu Feng, Ruixiao Shi, Jing Wang, Xin Geng, Yong Rui
KDiv organizes diffusion generators during pretraining by separating shared learngenes from condition-specific tailors, whose contributions are continuously routed from class, text, or image conditions. For text-conditioned generation, it reports average LPIPS 0.519, SSIM 0.450, and CLIP-T 25.20, compared with LoRA’s 0.544, 0.422, and 24.52 across eight datasets. These averages do not guarantee better results on every target.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.