The Lab · instrument 08/09 · interactive

Steer the denoiser

Four images denoise side by side under the real classifier-free-guidance formula — turn one knob and watch literalness, contrast clipping, and diversity collapse trade against each other.

What's real, what's toy

The formula is the genuine article from diffusion models: at every step the sampler moves along e_u + g·(e_c − e_u) — the unconditioned direction plus g times the difference the prompt makes. What's toy: instead of a trained network estimating those directions, an oracle knows each sample's destinations exactly. That removes the learned magic and leaves the steering geometry — which is precisely the part the knob controls in a real system.

Where to push

Set g to 0 and resample: the prompt contributes nothing, and you get four unrelated color dreams — the model's "unconditioned distribution." Set g to 1: four images that all satisfy the prompt differently — on target, still diverse. Now push past 2.5 and watch two things happen at once. Pixels extrapolate beyond the prompt's target and clip against the displayable range — that's the blown saturation and crunchy contrast every practitioner recognizes as over-guidance. And the four samples stop looking like four ideas: the shared prompt direction dominates whatever made each sample itself. Then try steps: at 5, the schedule can't shed its noise and the results stay gritty; by 30, smooth — compute buying cleanliness, the same dial production samplers price out.

Failure mode

Cranking guidance to fix a bad image — the exact reflex the diffusion entry warns about, now available to your own hands. When a sample disappoints at g = 1, raising g doesn't buy a better idea; it buys a louder version of the same one, with clipped highlights as the receipt. Resample first, adjust the prompt second, and move the guidance knob last, in small steps — in here, and out there.