Karras Sigma Schedule

1. Background Knowledge

In diffusion models, the forward process gradually adds Gaussian noise to a data sample over a sequence of timesteps. Traditional implementations (e.g., DDPM) use a discrete schedule of variance parameters $\beta_t$, but the EDM (Elucidated Diffusion Models) framework by Karras et al. (2022) reparameterizes the process using a continuous noise level $\sigma$. This simplifies the math and makes the sampling procedure more stable and interpretable.

The Karras schedule does not place sampling steps uniformly in $\sigma$-space. Instead, it applies a power warp controlled by the exponent $\rho$. By working in the transformed space $\sigma^{1/\rho}$, the schedule can concentrate more steps near the low-noise end (close to the clean data) when $\rho > 1$. This is critical because the final denoising steps require finer resolution to produce high-quality samples. The formula interpolates linearly between $\sigma_{\max}^{1/\rho}$ and $\sigma_{\min}^{1/\rho}$, then maps back to $\sigma$-space by raising the result to the power $\rho$.

A key practical detail is that the sampling loop requires a final step where $\sigma = 0$, representing the transition from the noisiest remaining state to a fully clean sample. This zero is appended after the $N$ computed warped sigmas.

2. Algorithm Approach

This is a direct formula evaluation problem. There is no iterative search, dynamic programming, or optimization involved. The approach is:

1. Compute the warped endpoints: $\sigma_{\max}^{1/\rho}$ and $\sigma_{\min}^{1/\rho}$.
2. For each index $i$ from $0$ to $N-1$, compute the linear interpolation coefficient $t_i = \frac{i}{N-1}$.
3. Evaluate the warped value: $w_i = \sigma_{\max}^{1/\rho} + t_i (\sigma_{\min}^{1/\rho} - \sigma_{\max}^{1/\rho})$.
4. Map back to $\sigma$-space: $\sigma_i = w_i^{\rho}$.
5. Append $0.0$ to the list.
6. Round all values to 4 decimal places.

The core pattern is vectorized or loop-based evaluation of a closed-form expression.

3. Step-by-Step Strategy

• Validate inputs: Ensure $N \geq 2$, $0 < \sigma_{\min} < \sigma_{\max}$, and $\rho > 0$.
• Precompute constants: Calculate sigma_max_pow = sigma_max ** (1.0 / rho) and sigma_min_pow = sigma_min ** (1.0 / rho).
• Iterate over indices: Loop $i$ from $0$ to $N-1$.
• Compute the interpolation fraction: t = i / (N - 1).
• Compute the warped intermediate value: w = sigma_max_pow + t * (sigma_min_pow - sigma_max_pow).
• Compute the final sigma: sigma = w ** rho.
• Append sigma to a result list.
• Append zero: Add 0.0 to the end of the list.
• Round: Apply round(x, 4) to every element in the list.
• Return: Return the rounded list.

Example skeleton:

def karras_sigmas(N, sigma_min, sigma_max, rho=7.0):
 sigmas = []
 smax_p = sigma_max ** (1.0 / rho)
 smin_p = sigma_min ** (1.0 / rho)
 for i in range(N):
 t = i / (N - 1)
 w = smax_p + t * (smin_p - smax_p)
 sigmas.append(w ** rho)
 sigmas.append(0.0)
 return [round(s, 4) for s in sigmas]

4. Common Pitfalls

• Off-by-one in interpolation: The denominator must be $N-1$, not $N$, so that $i=0$ gives exactly $\sigma_{\max}$ and $i=N-1$ gives exactly $\sigma_{\min}$. Using $N$ would shift the entire schedule.
• Forgetting the final zero: The output must have length $N+1$. Omitting the trailing 0.0 is a frequent error.
• Rounding precision: Round to 4 decimal places as specified. Rounding to 2 or 6 will fail test cases.
• Integer division: In Python 3, / performs float division, but be cautious if porting to other languages. Ensure i / (N - 1) produces a float.
• Negative base for fractional power: Since $\sigma_{\min} > 0$ and $\sigma_{\max} > 0$, the warped values $w_i$ remain positive, so $w_i^{\rho}$ is well-defined. However, if inputs were invalid (e.g., $\sigma_{\min} \leq 0$), fractional powers could produce complex numbers or errors.

5. Time & Space Complexity

• Time Complexity: $O(N)$, since we perform a constant amount of arithmetic for each of the $N$ steps.
• Space Complexity: $O(N)$, to store the list of $N+1$ sigmas.