ancestral sampling

**Ancestral sampling** is the **stochastic reverse diffusion method that samples new noise at each step using predicted mean and variance** - it follows the probabilistic reverse process and naturally supports output diversity. **What Is Ancestral sampling?** - **Definition**: Each reverse step draws from a conditional Gaussian distribution instead of a deterministic update. - **Noise Injection**: Fresh randomness is introduced repeatedly as the sample denoises. - **Model Dependency**: Uses network predictions for denoised direction plus variance parameterization. - **Trajectory Behavior**: Different random draws produce varied samples from the same prompt and seed space. **Why Ancestral sampling Matters** - **Diversity**: Stochasticity improves mode coverage and creative variation. - **Probabilistic Fidelity**: Matches the intended generative process in many diffusion formulations. - **Uncertainty Modeling**: Represents ambiguity in conditional generation tasks. - **Benchmark Use**: Common reference method for evaluating accelerated alternatives. - **Latency Cost**: Usually requires many steps and can be slower than ODE solvers. **How It Is Used in Practice** - **Variance Control**: Tune temperature or variance scaling to prevent excessive noise artifacts. - **Seed Strategy**: Generate multiple seeds for candidate selection in user-facing systems. - **Guidance Balance**: Avoid overly aggressive guidance that collapses stochastic diversity benefits. Ancestral sampling is **the canonical stochastic path for reverse diffusion generation** - ancestral sampling is preferred when diversity and probabilistic behavior matter more than minimum latency.