The Lab · instrument 08/16 · interactive

Steer the denoiser

Four images denoise side by side under the real classifier-free-guidance formula — turn one knob and watch literalness rise, channels clip, and four ideas take on one overdriven look.

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: the prompt contributes nothing, and you get four unrelated color dreams — the model's "unconditioned distribution." Set g to 1: four images that each satisfy the prompt with a different layout — on target, still varied. Now push past 2.5 and watch pixels extrapolate beyond the prompt's target and clip against the displayable range — the blown saturation and crunchy contrast every practitioner recognizes as over-guidance. The four stop looking like four different pictures not because guidance erased their differences but because so many channels have clipped that everything arrives blown out the same way. (In a real sampler all four would share one prompt, and high g genuinely collapses diversity too — here you're watching the clipping half of that story.) Then try steps: at 5 the image snaps to its target in a few violent jumps; at 30 you watch the noise schedule anneal it down gradually — in this oracle toy the step count buys you the view, while in a real sampler it buys the cleanliness as well.

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.