Reproduce “VCT: Training Consistency Models with Variational Noise Coupling” 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
Gianluigi Silvestri, Luca Ambrogioni, Chieh-Hsin Lai, Yuhta Takida, Yuki Mitsufuji
Researchers at Sony AI and academic institutions developed Variational Consistency Training (VCT), a method that stabilizes non-distillation Consistency Training by adaptively learning noise-data pairings. This approach results in improved image generation quality, including achieving a competitive FID of 2.02 for 2-step sampling on CIFAR-10 among non-distillation Consistency Models.
Key insights
- Adaptive learning of noise-data pairings (Variational Coupling) can significantly reduce training variance and stabilize direct Consistency Training.
- Framing this learned coupling within the Flow Matching framework and drawing parallels to Variational Autoencoders provides a strong theoretical foundation.
- Allowing the model to implicitly "move the trajectories towards the space of constancy" helps ameliorate problematic singularities in the Ordinary Differential Equation (ODE) flow.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.