Classifier Gradient Guidance Step

mediumPROGuidance and Latent DiffusionDiffusion ModelsDiffusion Models 1

Problem Statement

Before classifier-free guidance, Dhariwal & Nichol used an external classifier's gradient to nudge the noise prediction. Implement that guided-noise update.

Background

Classifier guidance shifts the predicted noise by the gradient of the classifier's log-probability of the target class, scaled by the noise level and a strength s:

$\hat{\varepsilon} = \varepsilon - s\,\sqrt{1 - \bar{\alpha}_t}\; \nabla_{x_t} \log p_\phi(y \mid x_t)$

The sqrt(1 - alpha_bar_t) factor converts a score shift into a noise shift. The gradient is supplied.

Your Task

Implement:

def classifier_guidance(eps, grad_log_p, alpha_bar_t, s):

Return the guided noise as a list rounded to 4 decimals.

Input Format

eps, grad_log_p: lists of equal length D.
alpha_bar_t (float) in (0, 1), s (float): guidance strength.

Output Format

A list of D floats rounded to 4 decimals.

Sample

print(classifier_guidance([0.0, 0.0], [1.0, -1.0], 0.75, 2.0))

Output:

[-1.0, 1.0]

Example

Input

print(classifier_guidance([0.0, 0.0], [1.0, -1.0], 0.75, 2.0))

Output

[-1.0, 1.0]

Reasoning

Calculate the noise scaling factor derived from the time step, which converts the score shift into a noise shift: $\sqrt{1 - \alpha_{\text{bar}}_t} = \sqrt{1 - 0.75} = \sqrt{0.25} = 0.5$ .
Determine the effective guidance magnitude by multiplying the strength $s$ by the scaling factor: $2.0 \times 0.5 = 1.0$ .
Compute the guided noise for the first dimension by subtracting the scaled gradient from the predicted noise: $0.0 - (1.0 \times 1.0) = -1.0$ .
Compute the guided noise for the second dimension by subtracting the scaled gradient from the predicted noise: $0.0 - (1.0 \times -1.0) = 1.0$ .
The final output is [-1.0, 1.0]

Constraints

len(eps) == len(grad_log_p), 0 < alpha_bar_t < 1.
eps_guided = eps - s*sqrt(1-alpha_bar_t)*grad.
Round to 4 decimals; avoid -0.0.

Learn this in

Conditional GenerationDiffusion 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 Negative Prompt as the Unconditional Branchmedium Latent Encode-Decode Round Tripmedium

Concept primerPREVIEW

1. Background Knowledge

In classifier-guided diffusion, a separate neural network (the classifier) is trained to predict the probability of a class label $y$ given a noisy latent $x_t$ . During sampling, we want the diffusion process to move toward regions of high $p_\phi(y \mid x_t)$ . The direction of this movement is given by the gradient $\nabla_{x_t} \log p_\phi(y \mid x_t)$ , which points toward higher log-probability for the target class.

The raw diffusion model predicts noise $\varepsilon$ at timestep $t$ . To incorporate classifier guidance, we shift this prediction by a scaled gradient term. The scaling factor $\sqrt{1 - \bar{\alpha}_t}$ is critical: it converts a score-space shift (in the $x_t$ domain) into an equivalent noise-space shift. This arises from the relationship between the score function $\nabla_{x_t} \log p_t(x_t)$ and the noise prediction $\varepsilon$ in the forward diffusion process, where $\bar{\alpha}_t$ is the cumulative product of per-step signal retention factors.

The guidance strength $s$ controls how aggressively the classifier steers the generation. When $s = 0$ , the classifier has no effect and the output equals the original noise prediction. As $s$ increases, the generated samples become more class-conditional but may lose diversity or fidelity.

2. Algorithm Approach

This is a pointwise element-wise transformation problem. For each dimension $i$ of the input vectors, you compute:

$\hat{\varepsilon}_i = \varepsilon_i - s \cdot \sqrt{1 - \bar{\alpha}_t} \cdot \nabla_i$

where $\nabla_i$ is the $i$ -th component of the gradient vector. The entire operation reduces to:

Compute the scalar coefficient $c = s \cdot \sqrt{1 - \bar{\alpha}_t}$ .
For each index $i$ , subtract $c \cdot \texttt{grad\_log\_p}[i]$ from $\texttt{eps}[i]$ .
Round each result to 4 decimal places.

No loops over timesteps or batch dimensions are needed—this is a single-step, single-sample update.

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.

PRO

Editor locked

The code editor is locked for Pro problems. It is only available for free problems. Please upgrade to gain access to the code editor for all problems.

solution.py

Python 3.13

Test results