CFG Guidance Scale, Playable

EP 01

The trick — run the model twice, once deaf

At every denoising step the model actually runs twice: once seeing your prompt ("a red fox"), once with the prompt blanked out. Two predictions of where the image should head. The difference between them is a pure, distilled arrow meaning "this is what the prompt specifically wants." Step through a denoise below and watch the two predictions and their difference — the whole of CFG is that gold arrow.

✓ good case: the two runs disagree a lot → strong prompt signal to amplify✗ bad case: vague prompt ("nice picture") — the two runs barely differ, the arrow is noise, guidance has nothing to grab

—

Vectors shown are 2D stand-ins for the real high-dimensional noise predictions; their measured angle-gap drives everything downstream, exactly as in the real algorithm: guided = unconditional + scale × (conditional − unconditional).

"Classifier-free" means the model is its own critic: instead of a separate classifier network pushing images toward "fox-ness" (the old way), the model's deaf run serves as the baseline to subtract. One clean subtraction replaced an entire second model — which is why every diffusion system adopted it within a year.

EP 02

The arrow — extrapolate past the prompt

Guidance scale is just how far you slide along that arrow. Scale 1 = trust the conditional prediction as-is. Scale 7 = go seven times farther in the prompt's direction than the model itself predicted — deliberate exaggeration, past the data, into caricature country. Drag the dial and watch the output point slide along the line, off the manifold of images the model was trained on.

✓ good case: scale 5-8 — safely exaggerated, prompt features amplified, still on plausible ground✗ bad case: scale 25 — you've extrapolated into a region no training image ever occupied; the math happily goes there, the picture doesn't survive the trip

guidance scale 7.0 —

The grey cloud is the data manifold (where real training images live); the readout measures the guided point's distance from it. CFG is literally linear extrapolation — nothing stops it from leaving the map.

Guidance trades diversity and realism for obedience. Formally it sharpens the distribution toward high prompt-likelihood — which is why guided samples are more "on prompt" but also more same-y, more saturated, more caricatured. The dial isn't quality; it's exaggeration.

EP 03

Burn one yourself — from mush to poster to charcoal

The famous sweep, live. The canvas generates an abstract "image" via actual guided denoising of a 2D color field — the same math at toy scale. Drag from 0 to 30 and watch the three regimes: mush (low: pretty but ignores the prompt's palette/structure), the sweet spot (5-9), and deep-fry (20+: colors clip to maximum saturation, contrast crushes, detail dies). The saturation-clipping meter counts the actual burned pixels.

✓ good case: scale 7 — structured, saturated-but-alive, 0.4% clipped pixels✗ bad case: scale 25 — 38% of pixels clipped to pure max; the "obedient" image is a charcoal briquette

guidance scale 7 —

Real guided denoising on a 48×32 color field: conditional prediction pulls toward the "prompt" (a warm structured pattern), unconditional toward generic smoothness; clipping percentage is counted from the actual pixel values.

Deep-frying isn't a bug in one app — it's the arithmetic of over-extrapolation: values pushed past the representable range clip, and clipping IS the burnt look. Every "why do my images look microwaved" thread ends at this dial, and every fix (dynamic thresholding, rescaled CFG) is a way of extrapolating without clipping.

EP 04

The 2× tax — and the distillation heist

That "run twice, once deaf" trick has a price: every denoising step costs double. 50 steps become 100 forward passes. The meter below does the honest accounting for your settings — then shows the heist that modern fast models pulled: guidance distillation, where a student model is trained to directly output the guided prediction in one pass, baking the dial's effect into the weights.

✗ bad case: 50 steps × CFG = 100 passes per image — half your GPU bill is the deaf twin✓ good case: distilled model — guided quality, single pass per step, dial baked in (the reason turbo/lightning models don't expose a CFG slider)

denoise steps 50 —

Pass counts are exact; the throughput panel assumes a fixed per-pass cost. Distillation's catch, shown honestly: the dial is frozen at training time — you lose the knob to gain the speed.

Follow the compute and CFG explains a product mystery: why fast image models (SDXL-Turbo, Lightning, Flux-schnell) often lack a guidance slider entirely. They didn't remove the feature — they fossilized it. The obedience is still there; the dial was paid for in advance.

EP 05

One number is too crude — schedule the dial

The frontier refinement: constant guidance is wrong at both ends. Early steps decide composition (guidance helps a lot); late steps polish texture (guidance mostly burns). So schedule it: high early, low late. Design your own curve below and read the two measured meters — prompt adherence and burn damage. The constant-7 baseline is marked; beat it.

✓ good case: 12 → 3 declining schedule — adherence of constant-10, burn of constant-4✗ bad case: inverted schedule (low early, high late) — weak composition AND fried texture, the worst of both ends

early guidance 12 late guidance 3 —

Adherence is dominated by early-step guidance, burn by late-step guidance — the interpolated schedule's scores are computed from that (empirically observed) split, with the constant-7 reference measured the same way.

Every mature knob eventually becomes a curve. Guidance intervals, rescaled CFG, per-step schedules — the 2023-26 refinements all say the same thing: obedience belongs at the composition stage, freedom at the detail stage. If your generator has an "advanced" tab, this is what's hiding in it.