Learning Where to Steer Noise
Generative models for offline multi-objective optimization usually steer every sampling step. We steer only the starting noise, along the few directions the objectives actually care about.
Research, one denoising step at a time: clean explanations of a noisy world.
Generative models for offline multi-objective optimization usually steer every sampling step. We steer only the starting noise, along the few directions the objectives actually care about.
Multi-objective combinatorial optimization as online learning over decomposed decision spaces: 80–98% of specialized solvers at a tiny fraction of the cost.
Temporally aligned audio–video generation with latent diffusion models, from my Alexa+ AI internship.