Reproduce “Coevolving Representations in Joint Image-Feature Diffusion” 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
Theodoros Kouzelis, Spyros Gidaris, Nikos Komodakis
Coevolving Representation Diffusion (CoReDi) enables the semantic representation space in joint image-feature diffusion models to adapt dynamically during training, rather than relying on fixed projections. This method achieved up to 2x faster convergence while matching or surpassing baseline image generation quality, such as an FID of 1.58 with half the training steps compared to leading methods, by ensuring the representations develop stronger spatial organization.
Key insights
- Dynamically adapting the semantic representation space during training (coevolution) significantly enhances generative model performance, leading to faster convergence and higher sample quality.
- Stable optimization of coevolving representations requires careful architectural and training modifications, particularly preventing trivial loss minimization and feature collapse.
- The coevolving representations develop a stronger, more beneficial spatial organization over time, which directly contributes to improved image synthesis quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.