heun method sampling

**Heun method sampling** is the **second-order predictor-corrector integration method that refines Euler updates for more accurate diffusion trajectories** - it improves stability and fidelity with modest extra computation. **What Is Heun method sampling?** - **Definition**: Computes a predictor step then corrects with an averaged derivative estimate. - **Order Advantage**: Second-order accuracy reduces integration error at fixed step counts. - **Cost Profile**: Requires additional evaluations but usually remains efficient in practice. - **Use Context**: Common choice when quality must improve without jumping to complex multistep solvers. **Why Heun method sampling Matters** - **Quality Gain**: Often yields cleaner detail and fewer trajectory artifacts than Euler. - **Stability**: Better handles stiff regions in guided sampling dynamics. - **Balanced Tradeoff**: Moderate overhead for meaningful visual improvements. - **Production Utility**: Suitable for balanced latency-quality presets in serving systems. - **Tuning Need**: Still depends on timestep spacing and model parameterization quality. **How It Is Used in Practice** - **Preset Design**: Use Heun for mid-latency modes where Euler quality is insufficient. - **Grid Optimization**: Test step spacings jointly with guidance scales and seed diversity. - **Fallback Logic**: Retain Euler fallback for edge-case numerical failures in rare prompts. Heun method sampling is **a strong second-order sampler for balanced diffusion inference** - Heun method sampling is a practical upgrade path when teams need better quality without major complexity.