Reproduce “Reflected Schrödinger Bridge Matching” 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
Marcus Häggbom, Viktor Nilsson, Pierre Nyquist, Joakim andén
α-RSBM adapts bridge matching to image translation with pixel values constrained to the unit cube. It uses reflected Brownian paths that bounce off boundaries, plus a reflected-score target and sampled bridge midpoints, avoiding full-path training. Experiments report similar visual quality to non-reflected models while keeping outputs in range; results are limited to a driftless reflected Brownian reference on the unit cube.
Key insights
- The core mathematical and algorithmic principles of efficient unconstrained Schrödinger bridge training (α-IMF) can be rigorously extended to incorporate reflected dynamics, providing a principled method for domain confinement.
- By restricting the reference process to reflected scaled Brownian motion and the domain to the unit hypercube, the necessary score functions and reflected bridge sampling can be computed efficiently, making high-dimensional reflected SB training feasible.
- Integrating reflected dynamics intrinsically eliminates out-of-bounds data values, addressing a fundamental limitation of many generative models without the need for post-processing or a significant impact on generative performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.