DDIM with Stochasticity (general eta)

mediumPRODDPM and DDIM SamplingDiffusion ModelsDiffusion Models 1

Problem Statement

Implement the general DDIM update with the interpolation parameter eta sweeping from deterministic (eta=0) to full DDPM (eta=1).

Background

$\sigma_t = \eta \sqrt{\frac{1-\bar{\alpha}_{t'}}{1-\bar{\alpha}_t}}\sqrt{1 - \frac{\bar{\alpha}_t}{\bar{\alpha}_{t'}}}$

$\hat{x}_0 = \frac{x_t - \sqrt{1-\bar{\alpha}_t}\,\hat{\varepsilon}}{\sqrt{\bar{\alpha}_t}}, \qquad x_{t'} = \sqrt{\bar{\alpha}_{t'}}\,\hat{x}_0 + \sqrt{1-\bar{\alpha}_{t'}-\sigma_t^2}\;\hat{\varepsilon} + \sigma_t z$

Clamp the direction-term square root argument at 0. Add sigma_t * z only when eta > 0.

Your Task

Implement:

def ddim_step(x_t, eps, alpha_bar_t, alpha_bar_prev, eta=0.0, z=None):

Return x_{t'} as a list rounded to 4 decimals.

Input Format

x_t, eps: lists of length D; z: list or None.
alpha_bar_t, alpha_bar_prev (float); eta (float).

Output Format

A list of D floats rounded to 4 decimals.

Sample

print(ddim_step([0.4, -0.6], [0.1, 0.2], 0.5, 0.8, 0.0))

Output:

[0.4612, -0.8484]

Example

Input

print(ddim_step([0.4, -0.6], [0.1, 0.2], 0.5, 0.8, 0.0))

Output

[0.4612, -0.8484]

Reasoning

Compute the noise scale $\sigma_t$ : Since $\eta = 0.0$ , the stochastic term vanishes, so $\sigma_t = 0$ .
Estimate the initial state $\hat{x}_0$ by removing the noise component from $x_t$ : Using $\sqrt{1 - \bar{\alpha}_t} = \sqrt{0.5} \approx 0.7071$ and $\sqrt{\bar{\alpha}_t} = \sqrt{0.5} \approx 0.7071$ , we get $\hat{x}_0 = \frac{x_t - 0.7071 \cdot \hat{\varepsilon}}{0.7071}$ . For the first element: $\frac{0.4 - 0.7071(0.1)}{0.7071} \approx 0.4596$ ; for the second: $\frac{-0.6 - 0.7071(0.2)}{0.7071} \approx -0.8485$ .
Calculate the deterministic direction term: The coefficient is $\sqrt{1 - \bar{\alpha}_{t'} - \sigma_t^2} = \sqrt{1 - 0.8 - 0} = \sqrt{0.2} \approx 0.4472$ . Multiplying by $\hat{\varepsilon}$ gives $[0.0447, 0.0894]$ .
Combine the scaled prediction and direction term to find $x_{t'}$ $x_{t^{'}}$ : Multiply $\hat{x}_0$ $\overset{x}{^}_{0}$ by $\sqrt{\bar{\alpha}_{t'}} = \sqrt{0.8} \approx 0.8944$ $\overset{α}{ˉ}_{t^{'}} = 0.8 \approx 0.8944$ to get $[0.4111, -0.7589]$ $[0.4111, - 0.7589]$ , then add the direction term: $[0.4111 + 0.0447, -0.7589 + 0.0894] = [0.4558, -0.6695]$ $[0.4111 + 0.0447, - 0.7589 + 0.0894] = [0.4558, - 0.6695]$ . Correction: Re-evaluating the exact algebraic simplification for $\eta=0$ $η = 0$ where $x_{t'} = \sqrt{\bar{\alpha}_{t'}}\hat{x}_0 + \sqrt{1-\bar{\alpha}_{t'}}\hat{\varepsilon}$ $x_{t^{'}} = \overset{α}{ˉ}_{t^{'}} \overset{x}{^}_{0} + 1 - \overset{α}{ˉ}_{t^{'}} \overset{ε}{^}$ :
- Element 1: $\sqrt{0.8}\left(\frac{0.4 - \sqrt{0.5}(0.1)}{\sqrt{0.5}}\right) + \sqrt{0.2}(0.1) \approx 0.4612$
- Element 2: $\sqrt{0.8}\left(\frac{-0.6 - \sqrt{0.5}(0.2)}{\sqrt{0.5}}\right) + \sqrt{0.2}(0.2) \approx -0.8484$
The final output is [0.4612, -0.8484]

Constraints

len(x_t) == len(eps); alpha-bar values in (0, 1].
Clamp 1 - alpha_bar_prev - sigma^2 at 0 inside the square root.
Add sigma*z only when eta > 0; round to 4 decimals; avoid -0.0.

More from Diffusion Models 1

Classifier-Free Guidance: Uncond/Cond Mixingeasy DDPM Posterior Mean and Variance from x0medium Adaptive CFG Scheduling over the Trajectoryhard Classifier Gradient Guidance Stepmedium Negative Prompt as the Unconditional Branchmedium

Concept primerPREVIEW

1. Background Knowledge

DDPM (Denoising Diffusion Probabilistic Models) defines a forward noising process where Gaussian noise is gradually added to data, and a reverse denoising process where a neural network predicts the noise $\hat{\varepsilon}$ to recover the clean signal. The standard DDPM reverse step is stochastic and follows a specific Gaussian transition kernel.

DDIM (Denoising Diffusion Implicit Models) generalizes the reverse process by introducing a deterministic path. When the interpolation parameter $\eta = 0$ , the update becomes fully deterministic, meaning the same noise prediction $\hat{\varepsilon}$ always leads to the same $x_{t'}$ . This allows for faster sampling (fewer steps) and enables techniques like classifier guidance.

The general DDIM update bridges these two extremes. By varying $\eta$ from $0$ to $1$ , you interpolate between the deterministic DDIM path and the stochastic DDPM path. The key insight is that the "noise" added at each step is controlled by $\sigma_t$ , which depends on $\eta$ and the ratio of cumulative noise schedules ( $\bar{\alpha}$ ). When $\eta=1$ , $\sigma_t$ recovers the standard DDPM noise level.

2. Algorithm Approach

The implementation follows a direct formulaic approach:

Compute $\sigma_t$ : Use the provided formula involving $\eta$ , $\bar{\alpha}_t$ , and $\bar{\alpha}_{t'}$ (which is alpha_bar_prev in the function signature).
Predict $\hat{x}_0$ : Use the current noisy sample $x_t$ and the predicted noise $\hat{\varepsilon}$ to estimate the clean data.
Reconstruct $x_{t'}$ : Combine the predicted clean data $\hat{x}_0$ , the predicted noise $\hat{\varepsilon}$ , and the stochastic term $\sigma_t z$ using the specific coefficients derived from the DDIM derivation.
Handle Edge Cases: Ensure numerical stability by clamping square root arguments and conditionally adding the stochastic term.

3. Step-by-Step Strategy

Calculate $\sigma_t$ :

Compute the term inside the first square root: $\frac{1-\bar{\alpha}_{t'}}{1-\bar{\alpha}_t}$ .
Compute the term inside the second square root: $1 - \frac{\bar{\alpha}_t}{\bar{\alpha}_{t'}}$ .
Multiply these by $\eta^2$ (since $\sigma_t = \eta \sqrt{\dots}$ , so $\sigma_t^2 = \eta^2 \dots$ ).
Take the square root to get $\sigma_t$ .

PRO

Continue the full explanation

You're reading the free preview. Unlock the complete walkthrough, the code editor, test runner and reference solution with Premium.