NoiseRater: Meta-Learned Noise Valuation for Diffusion Model Training
Researchers have introduced NoiseRater, a novel meta-learning framework designed to enhance the training efficiency and generation quality of diffusion models. Challenging the conventional assumption that all injected noise is uniformly informative, this study proposes a parametric noise rater that assigns instance-level importance scores to individual noise realizations based on data and timestep. This mechanism enables adaptive reweighting of the training objective. The rater is optimized via bilevel optimization to improve downstream validation performance following inner-loop diffusion updates. To facilitate practical deployment, the authors developed a decoupled two-stage pipeline, transitioning from soft weighting during meta-training to hard noise selection in standard training phases. Extensive experiments conducted on FFHQ and ImageNet datasets demonstrate that prioritizing informative noise samples significantly boosts both training speed and output quality. These findings establish noise valuation as a critical, previously underexplored dimension for optimizing diffusion model training. The research highlights that not all noise contributions are equal, offering a new axis for improvement in generative AI tasks. The associated code has been made publicly available to support further community exploration and implementation of this methodology.
Wire timeline
NoiseRater: Meta-Learned Noise Valuation for Diffusion Model Training
Researchers have introduced NoiseRater, a novel meta-learning framework designed to enhance the training efficiency and generation quality of diffusion models. Challenging the conventional assumption that all injected noise is uniformly informative, this study proposes a parametric noise rater that assigns instance-level importance scores to individual noise realizations based on data and timestep. This mechanism enables adaptive reweighting of the training objective. The rater is optimized via bilevel optimization to improve downstream validation performance following inner-loop diffusion updates. To facilitate practical deployment, the authors developed a decoupled two-stage pipeline, transitioning from soft weighting during meta-training to hard noise selection in standard training phases. Extensive experiments conducted on FFHQ and ImageNet datasets demonstrate that prioritizing informative noise samples significantly boosts both training speed and output quality. These findings establish noise valuation as a critical, previously underexplored dimension for optimizing diffusion model training. The research highlights that not all noise contributions are equal, offering a new axis for improvement in generative AI tasks. The associated code has been made publicly available to support further community exploration and implementation of this methodology.
cs.AI updates on arXiv.org