Adaptive CFG Scheduling over the Trajectory

hardPROGuidance and Latent DiffusionDiffusion ModelsDiffusion Models 1

Problem Statement

A single fixed guidance scale is suboptimal: strong guidance early adds detail, strong guidance late over-saturates. Implement a linearly interpolated CFG schedule and apply it across a batch of per-step noise predictions.

Background

Given K sampling steps ordered from first (high noise) to last (low noise), the guidance scale ramps linearly from w_start at step 0 to w_end at step K-1:

$w_k = w_{\text{start}} + \frac{k}{K-1}\,(w_{\text{end}} - w_{\text{start}})$

(with w_0 = w_start when K == 1). At each step apply CFG with that step's scale:

$\hat{\varepsilon}_k = \varepsilon^{\text{uncond}}_k + w_k\,(\varepsilon^{\text{cond}}_k - \varepsilon^{\text{uncond}}_k)$

Your Task

Implement:

def scheduled_cfg(eps_uncond, eps_cond, w_start, w_end):

eps_uncond[k], eps_cond[k]: the per-step predictions (each a list of length D), K steps total.

Return the list of K guided predictions, each a list rounded to 4 decimals.

Input Format

eps_uncond, eps_cond: K x D nested lists.
w_start, w_end (float).

Output Format

A K x D nested list rounded to 4 decimals.

Sample

u = [[0.0], [0.0], [0.0]]
c = [[1.0], [1.0], [1.0]]
print(scheduled_cfg(u, c, 10.0, 2.0))

Output:

[[10.0], [6.0], [2.0]]

Example

Input

u = [[0.0], [0.0], [0.0]]
c = [[1.0], [1.0], [1.0]]
print(scheduled_cfg(u, c, 10.0, 2.0))

Output

[[10.0], [6.0], [2.0]]

Reasoning

Determine the number of steps $K=3$ and compute the linearly interpolated guidance scales $w_k$ for $k \in \{0, 1, 2\}$ using $w_k = 10.0 + \frac{k}{2}(2.0 - 10.0)$ , yielding $w_0 = 10.0$ , $w_1 = 6.0$ , and $w_2 = 2.0$ .
Calculate the difference between conditional and unconditional predictions for each step: $\varepsilon^{\text{cond}}_k - \varepsilon^{\text{uncond}}_k = 1.0 - 0.0 = 1.0$ for all $k$ .
Apply the CFG formula $\hat{\varepsilon}_k = 0.0 + w_k(1.0)$ at step 0 to get $\hat{\varepsilon}_0 = 10.0 \cdot 1.0 = 10.0$ .
Apply the formula at step 1 to get $\hat{\varepsilon}_1 = 6.0 \cdot 1.0 = 6.0$ .
Apply the formula at step 2 to get $\hat{\varepsilon}_2 = 2.0 \cdot 1.0 = 2.0$ .
The final output is [[10.0], [6.0], [2.0]]

Constraints

len(eps_uncond) == len(eps_cond) == K, K >= 1; rows share length D.
w_k ramps linearly from w_start (k=0) to w_end (k=K-1); K==1 uses w_start.
Per step: eps_uncond + w_k*(eps_cond - eps_uncond); round to 4 decimals; avoid -0.0.

Learn this in

Sampling & AccelerationDiffusion Study Plan Latent Diffusion & Stable DiffusionDiffusion Study Plan Conditional GenerationDiffusion Study Plan

More from Diffusion Models 1

Classifier-Free Guidance: Uncond/Cond Mixingeasy DDPM Posterior Mean and Variance from x0medium Classifier Gradient Guidance Stepmedium Negative Prompt as the Unconditional Branchmedium Latent Encode-Decode Round Tripmedium

Concept primerPREVIEW

1. Background Knowledge

Classifier-Free Guidance (CFG) is a technique used in latent diffusion models to steer the denoising process toward the conditioning signal (e.g., a text prompt) without requiring a separate classifier. At each sampling step, the model makes two noise predictions: one conditioned on the prompt ( $\varepsilon^{\text{cond}}$ ) and one unconditioned ( $\varepsilon^{\text{uncond}}$ ). The guided prediction is a linear extrapolation:

$\hat{\varepsilon} = \varepsilon^{\text{uncond}} + w\,(\varepsilon^{\text{cond}} - \varepsilon^{\text{uncond}})$

where $w$ is the guidance scale. A larger $w$ pushes the sample further from the unconditioned manifold, increasing fidelity to the prompt but risking over-saturation or artifacts.

A single fixed $w$ is suboptimal because the noise level changes dramatically across the $K$ sampling steps. Early steps operate at high noise, where strong guidance helps establish coarse structure and detail. Late steps operate at low noise, where strong guidance over-saturates colors and textures. An adaptive schedule addresses this by interpolating $w$ linearly from $w_{\text{start}}$ (step 0) to $w_{\text{end}}$ (step $K-1$ ):

$w_k = w_{\text{start}} + \frac{k}{K-1}\,(w_{\text{end}} - w_{\text{start}})$

This is a simple linear interpolation (lerp) between two endpoints. When $K=1$ , the fraction $\frac{k}{K-1}$ is undefined, so the convention is $w_0 = w_{\text{start}}$ .

2. Algorithm Approach

The problem is a straightforward element-wise vector operation combined with a linear schedule computation:

Compute the schedule: For each step index $k \in \{0, \dots, K-1\}$ , calculate $w_k$ via linear interpolation.
Apply CFG per step: For each step $k$ , compute the guided prediction by combining the unconditioned and conditioned noise vectors using the scalar $w_k$ .
Round and return: Round each component to 4 decimal places and assemble the result.

No iterative refinement, sorting, or complex data structures are needed. The core pattern is a nested loop (or vectorized operation) over steps and dimensions.

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.