Learned noise schedule is a diffusion model technique where the noise addition schedule is optimized during training — rather than using fixed schedules like linear or cosine, the model learns optimal noise levels for each timestep.

What Is a Learned Noise Schedule?

• Definition: Neural network predicts optimal noise levels per timestep.
• Contrast: Fixed schedules (linear, cosine) use predetermined values.
• Benefit: Adapts to specific data distribution and model architecture.
• Training: Schedule parameters learned alongside denoiser.
• Result: Potentially faster convergence and better quality.

Why Learned Schedules Matter

• Data-Adaptive: Optimal schedule varies by image type.
• Quality: Can outperform hand-tuned schedules.
• Efficiency: Fewer steps needed with optimal schedule.
• Automation: No manual hyperparameter tuning.
• Research: Reveals insights about diffusion process.

Fixed vs Learned Schedules

Fixed (Linear, Cosine):

• Simple, well-understood.
• Works reasonably across domains.
• May not be optimal for specific tasks.

Learned:

• Adapts to data and architecture.
• More complex training.
• Can discover better schedules.

Examples

• EDM (Elucidating Diffusion Models): Learned schedule.
• Improved DDPM: Learned variance schedule.
• VDM (Variational Diffusion Models): End-to-end learned.

Learned noise schedules enable optimal diffusion training — adapting to your specific data and model.