Diffusion-free Score Matching for Exact Conditional Sampling
DISCO learns a single score field from slightly perturbed data, allowing the conditional score to be obtained exactly.
Conventional diffusion models learn scores across multiple noise levels, where conditioning information from clean data is unavailable. Thomas Wedenig and Robert Peharz’s method removes the need for those time-dependent fields. The authors report sample quality competitive with state-of-the-art diffusion models, a more faithful representation of the data distribution, and more accurate sampling from arbitrary conditionals than standard heuristic samplers.
ArXiv · AI/CL/LG's note
Conventional diffusion models learn scores across multiple noise levels, where conditioning information from clean data is unavailable. Thomas Wedenig and Robert Peharz’s method removes the need for those time-dependent fields. The authors report sample quality competitive with state-of-the-art diffusion models, a more faithful representation of the data distribution, and more accurate sampling from arbitrary conditionals than standard heuristic samplers.
ArXiv · AI/CL/LG's note
score 5