The idea in plain language.
Analyze how noise levels and sampling interact, then make training and generation more stable.
How it works
The paper examines failure modes in high-dimensional score-based learning and sampling. It adjusts the treatment of noise levels and sampling and uses an exponential moving average of model weights. These changes help the method generate larger images while retaining the basic approach of learning score functions and sampling through progressively lower noise levels.
What to keep in mind
The improvements are demonstrated on selected image datasets. High visual quality does not resolve issues such as training-data bias or faithful conditional generation.
Source: Improved Techniques for Training Score-Based Generative Models. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.