Core Concepts
Moving Average Sampling in Frequency domain (MASF) enhances stability in generative diffusion models by leveraging prior samples and frequency domain decomposition.
Abstract
The content introduces MASF, a technique to stabilize diffusion models by utilizing moving average in the frequency domain. It reinterprets denoising as model optimization and decomposes samples into frequency components. Extensive experiments show significant performance improvements in image generation tasks.
-
Abstract
- MASF introduces moving average in the frequency domain to enhance stability in diffusion models.
- Reinterprets denoising as model optimization and decomposes samples into frequency components.
- Extensive experiments demonstrate significant performance improvements in image generation tasks.
-
Introduction
- Diffusion models revolutionize generative tasks in computer vision.
- Forward process introduces noise, denoising network produces images.
- Compared to GANs, diffusion models offer training stability and high-quality samples.
-
Proposed Approach
- MASF reframes denoising as model optimization and utilizes moving average in the frequency domain.
- Decomposes samples into frequency components for dynamic evolution.
- Weighting scheme prioritizes low-frequency components early and high-frequency components later.
-
Experiments
- MASF improves FID scores in conditional and unconditional models on various datasets.
- Integration with other sampling techniques further enhances performance.
- Ablation studies show the importance of each component in MASF.
-
Conclusion
- MASF significantly improves stability in diffusion models for image generation.
- Negligible computational overhead and seamless integration into existing models.
Stats
MASF brings only 0.97% extra complexity cost.
FID scores consistently improve with the integration of MASF.
Linear weight in adaptive weighting scheme outperforms constant and quadratic weights.
Frequency weighting scheme with increasing weights for high-frequency components shows the best results.
Quotes
"MASF introduces negligible computational overhead and can be readily integrated into existing diffusion models."
"Extensive experiments validate that MASF significantly improves performance across various datasets, models, and sampling techniques."