Reproduce “ROMS-IMLE: A Minimalist Approach to Competitive Single-Step Generative Modelling” 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
Chirag Vashist, Ke Li
ROMS-IMLE, a single-step generative model, achieves competitive image generation quality and diversity by enhancing the Implicit Maximum Likelihood Estimation (IMLE) framework. It demonstrated an FID of 2.56 on ImageNet 256, utilizing 54% fewer parameters and 250 times fewer function evaluations compared to leading multi-step models.
Key insights
- Competitive high-quality image generation is achievable with a minimalist, single-step generative model, challenging the necessity of iterative sampling processes.
- The success of complex stochastic interpolant models may be largely attributed to 'per-stage supervision' across their compositional structure, rather than solely iterative refinement at inference.
- Enhancements to the IMLE framework, such as multi-stage supervision and robust loss functions, can significantly improve its performance to compete with state-of-the-art models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.