Score-Based Generative Modeling Framework
What you'll learn: How to formalize score-based generative modeling by systematically adding noise to data and learning time-dependent score functions to reverse the process.
The Core Idea
Score-based generative modeling brings together everything you've learned about score functions, noise conditioning, and Langevin dynamics into one unified framework. The key insight: instead of learning scores for a single noise level, you perturb data with progressively increasing noise levels and learn a time-dependent score function that knows how to denoise at each noise scale.
Think of it like teaching someone to restore old photographs. You don't just show them one badly damaged photo—you show them photos with slight blur, medium blur, heavy blur, and severe blur. They learn restoration strategies specific to each degradation level. Similarly, your model learns score functions conditioned on the noise level (or "time").
The Framework Components
- Data Perturbation: Add noise to clean data
x₀according to a schedule, creatingx_tat different noise levelst - Time-Dependent Score Network: Train a neural network
s_θ(x_t, t)that estimates the score∇log p_t(x_t)at noise levelt - Reverse Sampling: Use annealed Langevin dynamics to start from pure noise and gradually denoise by following the learned scores
The beauty is that by learning scores across all noise levels, you create a smooth path from random noise back to clean data. Each "timestep" has its own score function, and together they form a continuous denoising trajectory.
Key Takeaway: Score-based generative modeling formalizes generation as learning time-dependent score functions across multiple noise scales, enabling you to reverse a gradual noising process through principled sampling with Langevin dynamics.