linear noise schedule

**Linear noise schedule** is the **noise schedule where beta increases approximately linearly over diffusion timesteps** - it is simple to implement and historically common in early DDPM baselines. **What Is Linear noise schedule?** - **Definition**: Uses a straight-line interpolation between minimum and maximum noise variances. - **Behavior**: Often removes signal steadily but can over-degrade information in later timesteps. - **Historical Use**: Appears in foundational diffusion papers and many reference implementations. - **Compatibility**: Works with epsilon, x0, and velocity prediction objectives. **Why Linear noise schedule Matters** - **Reproducibility**: Simple formulation makes experiments easier to replicate across teams. - **Baseline Value**: Provides a consistent benchmark against newer schedule variants. - **Engineering Simplicity**: Requires minimal tuning to get a stable first training run. - **Known Limits**: Can be less efficient than cosine schedules in low-step sampling regimes. - **Decision Clarity**: Clear behavior helps diagnose schedule-related model failures. **How It Is Used in Practice** - **Initialization**: Start with standard beta ranges and verify gradient stability early in training. - **Comparison**: Benchmark against cosine schedule under identical solver and guidance settings. - **Retuning**: Adjust step count and guidance scale when switching from linear to alternative schedules. Linear noise schedule is **a dependable baseline schedule for diffusion experimentation** - linear noise schedule remains useful as a reference even when newer schedules outperform it.