Guidance: A cheat code for diffusion models
3 days ago
- #generative-models
- #diffusion-models
- #machine-learning
- Classifier-free diffusion guidance improves conditional diffusion model samples effectively and simply.
- Diffusion models model high-dimensional data distributions by predicting the score function.
- Conditional diffusion models use additional inputs to model conditional distributions.
- Classifier guidance uses a classifier to condition diffusion models but has limitations.
- Classifier-free guidance avoids training a separate classifier by using conditioning dropout.
- Guidance enhances conditioning signal adherence and sample quality at the cost of diversity.
- Guidance can be applied to autoregressive models but differs in effect compared to diffusion models.
- Temperature tuning in diffusion models is less effective than in autoregressive models.
- Guidance is a key factor in the recent success of diffusion models in image generation.