Noise Conditional Score Networks
What you'll learn: How conditioning score networks on noise level enables handling multiple scales of perturbation during diffusion model training.
The Core Challenge
When training diffusion models with denoising score matching, we add different amounts of noise to our data—from barely perceptible to completely overwhelming. A single neural network trying to estimate the score function for all these noise levels faces a difficult problem: the optimal denoising strategy changes drastically depending on how much noise is present.
Conditioning on Noise Level
Noise Conditional Score Networks solve this by explicitly telling the network how much noise is in the input. Instead of learning a single score function s(x), we learn a noise-level-aware score function s(x, σ) where σ represents the noise scale.
Think of it like asking someone to clean a window: if it's slightly dusty, you use a soft cloth; if it's heavily soiled, you need stronger tools. The cleaner needs to know the dirt level to choose the right approach.
How It Works
The network receives two inputs:
- The noisy data point
x - The noise level
σ(often embedded similarly to time steps)
During training, we:
- Sample noise levels from our schedule
- Perturb data with that specific noise level
- Train the network to predict the score for that noise level
This conditioning allows a single network to specialize its behavior across the entire range of noise scales, from heavily corrupted data (where large-scale structure matters) to nearly clean data (where fine details matter).
Key Takeaway: Noise Conditional Score Networks condition the score estimation network on the noise level σ, enabling a single model to handle multiple scales of perturbation effectively during training and sampling.