DDIM Deterministic Step (eta = 0)

easyPRODDPM and DDIM SamplingDiffusion ModelsDiffusion Models 1

Problem Statement

The deterministic DDIM update (eta = 0) is the workhorse of fast sampling. Implement one step given x_t, the predicted noise, and the two alpha-bar values.

Background

With eta = 0 there is no injected noise, so the update is

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

The same predicted noise is reused as the "direction pointing to x_t."

Your Task

Implement:

def ddim_deterministic(x_t, eps, alpha_bar_t, alpha_bar_prev):

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

Input Format

x_t, eps: lists of length D.
alpha_bar_t, alpha_bar_prev (float) in (0, 1].

Output Format

A list of D floats rounded to 4 decimals.

Sample

print(ddim_deterministic([1.4142, 0.0], [1.0, -1.0], 0.5, 1.0))

Output:

[1.0, 1.0]

Example

Input

print(ddim_deterministic([1.4142, 0.0], [1.0, -1.0], 0.5, 1.0))

Output

[1.0, 1.0]

Reasoning

Compute the predicted clean sample $\hat{x}_0$ using the formula $\hat{x}_0 = \frac{x_t - \sqrt{1 - \bar{\alpha}_t}\, \hat{\varepsilon}}{\sqrt{\bar{\alpha}_t}}$ . With $\bar{\alpha}_t = 0.5$ , the scaling factors are $\sqrt{1 - 0.5} = \sqrt{0.5} \approx 0.7071$ and $\sqrt{0.5} \approx 0.7071$ .
For the first dimension, substitute $x_t = 1.4142$ and $\hat{\varepsilon} = 1.0$ : $\hat{x}_{0,1} = \frac{1.4142 - (0.7071 \cdot 1.0)}{0.7071} = \frac{0.7071}{0.7071} = 1.0$ .
For the second dimension, substitute $x_t = 0.0$ and $\hat{\varepsilon} = -1.0$ : $\hat{x}_{0,2} = \frac{0.0 - (0.7071 \cdot -1.0)}{0.7071} = \frac{0.7071}{0.7071} = 1.0$ . Thus, $\hat{x}_0 = [1.0, 1.0]$ .
Compute the next state $x_{t'}$ using $x_{t'} = \sqrt{\bar{\alpha}_{t'}}\, \hat{x}_0 + \sqrt{1 - \bar{\alpha}_{t'}}\, \hat{\varepsilon}$ . Since $\bar{\alpha}_{t'} = 1.0$ , the coefficients are $\sqrt{1.0} = 1.0$ and $\sqrt{1 - 1.0} = 0.0$ .
Apply the coefficients to the vectors: $x_{t'} = 1.0 \cdot [1.0, 1.0] + 0.0 \cdot [1.0, -1.0] = [1.0, 1.0]$ .
The final output is [1.0, 1.0]

Constraints

len(x_t) == len(eps); alpha-bar values in (0, 1].
alpha_bar_prev = 1.0 yields exactly x0_hat (the direction term vanishes).
Round to 4 decimals; avoid -0.0.

Learn this in

Sampling & AccelerationDiffusion Study Plan

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) and DDIM (Denoising Diffusion Implicit Models) are two frameworks for sampling from a learned noise distribution. In DDPM, each reverse step is stochastic: you predict the noise $\hat{\varepsilon}$ , estimate the clean signal $\hat{x}_0$ , and then add fresh Gaussian noise to move to $x_{t-1}$ . DDIM generalizes this by introducing a parameter $\eta \in [0, 1]$ that controls how much stochasticity remains. When $\eta = 0$ , the process becomes deterministic: the same predicted noise $\hat{\varepsilon}$ is reused as the "direction" toward the noisy sample, and no new random noise is injected.

The forward process defines $\bar{\alpha}_t = \prod_{s=1}^{t} \alpha_s$ , where $\alpha_s = 1 - \beta_s$ . The noisy sample is $x_t = \sqrt{\bar{\alpha}_t}\, x_0 + \sqrt{1 - \bar{\alpha}_t}\, \varepsilon$ . The model predicts $\hat{\varepsilon}$ from $x_t$ . From this prediction you can recover an estimate of the clean signal:

$\hat{x}_0 = \frac{x_t - \sqrt{1 - \bar{\alpha}_t}\, \hat{\varepsilon}}{\sqrt{\bar{\alpha}_t}}$

For the deterministic DDIM step, you then reconstruct $x_{t'}$ (where $t' < t$ ) by blending $\hat{x}_0$ and $\hat{\varepsilon}$ using the previous alpha-bar value:

$x_{t'} = \sqrt{\bar{\alpha}_{t'}}\, \hat{x}_0 + \sqrt{1 - \bar{\alpha}_{t'}}\, \hat{\varepsilon}$

This is a linear combination, so the operation is element-wise over the feature dimension $D$ .

2. Algorithm Approach

The problem reduces to two vectorized (or element-wise) arithmetic operations:

Predict $\hat{x}_0$ : For each dimension $i$ , compute $\hat{x}_0[i] = \frac{x_t[i] - \sqrt{1 - \bar{\alpha}_t}\cdot \hat{\varepsilon}[i]}{\sqrt{\bar{\alpha}_t}}$ .
Reconstruct $x_{t'}$ : For each dimension $i$ , compute $x_{t'}[i] = \sqrt{\bar{\alpha}_{t'}}\cdot \hat{x}_0[i] + \sqrt{1 - \bar{\alpha}_{t'}}\cdot \hat{\varepsilon}[i]$ .

No loops over timesteps are needed—this is a single-step update. The key insight is that both $\sqrt{\bar{\alpha}_t}$ and $\sqrt{1-\bar{\alpha}_t}$ are scalars shared across all dimensions, so you compute them once and apply them element-wise.

3. Step-by-Step Strategy

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.