Ancestral Sampling

Ancestral sampling is an iterative process where each step is inherently stochastic. Starting from pure noise x_T ~ N(0, I), the model predicts a slightly denoised sample x_{t-1}. Crucially, this prediction is not deterministic; a new, small amount of Gaussian noise is added at each step according to the learned reverse process variance. This stochasticity is essential for generating diverse samples and exploring the data distribution.

• Process: x_T → x_{T-1} → ... → x_0
• Core Operation: x_{t-1} = μ_θ(x_t, t) + σ_t * z, where z ~ N(0, I)
• Outcome: A single, unique sample from the model's learned distribution.

The procedure strictly adheres to the reverse Markov chain defined by the trained model's parameters. Each step conditions only on the immediate previous state (x_t), not the full history. This Markov property is a direct inversion of the forward diffusion process and is encoded in the model's learned denoising function.

• Mathematical Basis: p_θ(x_{0:T}) = p(x_T) ∏{t=1}^{T} p_θ(x{t-1} | x_t)
• Conditional Distribution: p_θ(x_{t-1} | x_t) = N(x_{t-1}; μ_θ(x_t, t), Σ_θ(x_t, t))
• Contrast with DDIM: Unlike deterministic samplers (DDIM), it does not skip steps or use a non-Markovian process.

The amount of new noise injected at each denoising step is not arbitrary; it is governed by the model's learned variance schedule Σ_θ(x_t, t). In foundational models like DDPM, this is often fixed to a schedule derived from the forward process betas. However, the model can also learn to predict these variances, allowing the sampling process to adaptively control stochasticity.

• Common Practice (DDPM): Σ_θ(x_t, t) = σ_t² I, where σ_t² is a fixed scalar from the noise schedule.
• Learned Variances: More advanced models parameterize Σ_θ(x_t, t) with the neural network, potentially improving sample quality.
• Impact: The variance controls the trade-off between sample diversity and fidelity at each step.

Ancestral sampling in variance-preserving diffusion models is mathematically equivalent to sampling via Langevin dynamics using a learned score function. The model's noise prediction ϵ_θ(x_t, t) is directly proportional to the score (gradient of the log data density). Each sampling step can be viewed as moving towards higher data density (via the score) while adding noise for exploration.

• Equivalence: ϵ_θ(x_t, t) ∝ -score(x_t, t)
• Langevin Step: x_{t-1} = x_t + (α * score(x_t, t)) + (√(2α) * z)
• Foundation: This bridges Denoising Diffusion Probabilistic Models (DDPM) and Score Matching generative models.

The primary drawback of ancestral sampling is its computational expense. Generating a single sample requires performing a full forward pass of the neural network (e.g., a U-Net) at every timestep in the reverse chain. For high-fidelity image generation, this often means 50 to 1000 sequential evaluations, making it slow compared to single-step generators.

• Typical Steps: 50 (fast sampling) to 1000 (high quality) neural network evaluations.
• Bottleneck: Sequential dependency prevents parallelization across timesteps.
• Driver for Research: This cost motivates faster samplers (DDIM) and distillation methods (Consistency Models).

Ancestral sampling is the backbone for conditional generation techniques like Classifier-Free Guidance (CFG). In CFG, the core denoising step is modified. The model makes two predictions: one conditional (e.g., on a text prompt) and one unconditional. The final noise prediction is a weighted combination, steering the ancestral sampling trajectory towards regions of the data space that better match the condition.

• CFG Update: ϵ̃ = ϵ_uncond + guidance_scale * (ϵ_cond - ϵ_uncond)
• Process: This modified ϵ̃ is used in the standard ancestral sampling update rule.
• Result: Dramatically improved adherence to the input prompt or class label.

The foundational class of generative model that formalizes the framework for ancestral sampling. A DDPM learns to reverse a fixed forward process of gradually adding Gaussian noise to data. The model is trained to predict the noise at each timestep, enabling the iterative denoising procedure that defines ancestral sampling.

• Core Mechanism: Defines the Markov chain structure that ancestral sampling follows.
• Training Objective: Noise prediction, which is equivalent to learning the score function.
• Relation to Ancestral Sampling: Provides the learned reverse transition distributions p_θ(x_{t-1} | x_t) used in each sampling step.

The learned generative trajectory that ancestral sampling executes. Starting from pure noise x_T, this process applies a sequence of learned denoising steps to produce a data sample x_0. Each step is a stochastic transition defined by the model's parameters.

• Mathematical Definition: p_θ(x_{0:T}) = p(x_T) ∏_{t=1}^{T} p_θ(x_{t-1} | x_t)
• Stochasticity: In ancestral sampling, each reverse step injects new noise based on a learned variance schedule.
• Contrast with Deterministic Samplers: Unlike DDIM, the reverse process here is inherently Markovian and stochastic.

A predefined function β_t that controls the variance of the Gaussian noise added at each timestep t during the forward process. This schedule is critical because it determines the variances used in the ancestral sampling reverse steps.

• Key Parameters: Defines β_t (or α_t = 1 - β_t), which sets the noise level per step.
• Variance Types: Can be variance-preserving (total variance constant) or variance-exploding.
• Impact on Sampling: The learned reverse process variances in ancestral sampling are typically derived from this forward schedule.

An alternative training objective for generative models that is fundamentally connected to diffusion models. Instead of predicting noise, a score network is trained to estimate the score function—the gradient of the log data density.

• Score Function: ∇_x log p(x); points toward regions of higher data density.
• Equivalence: Under certain parameterizations, training a diffusion model's noise prediction network is equivalent to score matching.
• Sampling Link: Ancestral sampling in a trained model effectively follows the estimated score, with added noise for exploration.

Denoising Diffusion Implicit Models, a class of deterministic samplers that represent a key alternative to stochastic ancestral sampling. DDIMs redefine the forward process as non-Markovian, enabling faster generation with fewer steps.

• Deterministic vs. Stochastic: DDIM sampling follows an ODE, producing the same output for a given starting noise. Ancestral sampling is stochastic, producing varied outputs.
• Speed vs. Diversity: DDIM enables 10-50 step sampling versus 1000+ for ancestral sampling, but may trade off some mode coverage diversity.
• Same Model Weights: Both methods can use the same trained diffusion model backbone.

An ordinary differential equation derived from the continuous-time limit of a diffusion model's reverse-time SDE. It describes a deterministic trajectory that shares the same marginal distributions as the stochastic process used in ancestral sampling.

• Theoretical Bridge: Connects stochastic (ancestral) and deterministic (DDIM) sampling under one framework.
• Sampling Method: Solving this ODE with a numerical solver (like Euler) is a deterministic alternative to ancestral sampling.
• Flexibility: Allows for trade-offs between sampling speed and fidelity via choice of solver and step count.