What does classifier-free guidance (CFG) do?

In diffusion inference it combines the two predictions from each step by a guidance strength scale into that step's final prediction. The scale tunes how strongly the prompt is reflected.

How do you verify this combiner function?

Check its public promises in real numbers — that the result has the same shape as the inputs, that the defined behavior holds at boundary values like scale=1.0 and scale=0.0, and that a larger scale moves the output further toward cond.

Which libraries does this problem use?

Only Python and numpy. It does not involve an actual diffusion model, sampler, or scheduler — it focuses on the calculation that combines two numpy arrays.

combine classifier-free guidance

Your task

Diffusion inference makes two predictions per step — one with the prompt off (uncond) and one with it on (cond). It combines the two by a guidance strength scale to form the step's final prediction. The larger scale is, the more strongly the output is pulled toward cond. Complete apply_cfg(uncond, cond, scale). Example uncond = [0, 0, 0], cond = [1, 2, -1] → the larger scale is, the further the output moves toward cond.

Ground rules

•uncond and cond are real-valued numpy arrays of the same shape.
•scale is a real number ≥ 0.

Not required

•an actual diffusion model, sampler, or scheduler.
•variants such as negative prompts or a weighted sum of multiple conditions.

▶About this challenge

In diffusion inference each step yields two predictions — one with the prompt off (uncond) and one with it on (cond). This problem has you complete the combiner function in numpy, combining the two by a guidance strength scale into the step's final prediction. You work with same-shape real arrays and verify in code how the output should change as scale changes.

Classifier-free guidance (CFG) is the step in diffusion inference that combines two predictions by a single guidance strength to decide how strongly the prompt is reflected. Combining two arrays into one like this usually comes with promises (invariants) about what the result must be at particular scale values. So rather than trusting an output that merely looks plausible, the habit that matters is checking those promises in numbers. It is practice in numpy array arithmetic and in verifying boundary values (the two ends of scale).

What does classifier-free guidance (CFG) do?: In diffusion inference it combines the two predictions from each step by a guidance strength scale into that step's final prediction. The scale tunes how strongly the prompt is reflected.
How do you verify this combiner function?: Check its public promises in real numbers — that the result has the same shape as the inputs, that the defined behavior holds at boundary values like scale=1.0 and scale=0.0, and that a larger scale moves the output further toward cond.