CFG Scale & Guidance in AI Image Generation (2026)

If you have ever nudged the CFG slider from 7 to 15 and watched a clean portrait turn into an over-saturated, crispy-edged horror show, you already understand why guidance scale is the most misunderstood control in AI image generation. CFG scale governs the tension between what you asked for and what the model finds statistically natural — and in 2026, with distilled guidance, guidance rescaling, and CFG-free pipelines becoming standard, that tension is managed very differently than it was in the Stable Diffusion 1.5 era. This guide explains the mechanism, gives realistic starting values for current model families, and shares practical tuning strategies you can apply today.

What Is CFG Scale, Exactly?

CFG stands for Classifier-Free Guidance, introduced by Jonathan Ho and Tim Salimans in 2022. It is not a classifier in the traditional sense — no external model is scoring your prompt. Instead, the diffusion model runs two predictions at every step: one conditioned on your prompt, and one with the prompt removed (the “unconditional” or negative branch). The guidance scale controls how far the sampler extrapolates away from the unconditional prediction and toward the conditional one.

The 30-second version

The sampler’s effective prediction is essentially: unconditional + scale × (conditional − unconditional). A scale of 1.0 means pure conditional prediction, no amplification. A scale of 0 means your prompt is ignored entirely. Above 1.0, the difference vector gets exaggerated — the model leans harder into whatever features your prompt activated, including its worst habits.

CFG scale vs. “guidance scale”

In most 2026 interfaces these terms are interchangeable, but not always. Watch for:

• CFG / CFG scale — the classic two-pass classifier-free guidance value used in SD1.5, SDXL, and many SD3.5-era workflows.
• Guidance / distilled guidance — a separate conditioning parameter in Flux-style models, where CFG itself is often locked at 1.0 because guidance was baked into training.
• Prompt strength, match strength, text guidance — rebranded labels in various UIs that map to the same underlying concept.

This distinction matters enormously. Setting CFG to 7 on a model trained with distilled guidance will not make it follow your prompt better — it will usually just break the image.

What the Slider Actually Changes

Think of CFG as a contrast knob for prompt adherence, with hard diminishing returns.

• Too low (1–3 on classic models): soft, dreamy, low-contrast results with weak prompt adherence. Colors look washed out, detail is mushy, and composition drifts.
• The sweet spot: prompt elements appear reliably, lighting stays plausible, and textures retain nuance — skin pores rather than plastic sheen.
• Too high (12+ on classic models): crushed shadows, clipped highlights, halos, color banding, “deep-fried” saturation, repeated limbs, and hard-edged outlines. Paradoxically, adherence often gets worse because the model over-commits to isolated tokens.

The single most important diagnostic: if you raise CFG and the image gets more contrasty but no more accurate, you have passed the useful range.

Starting Values by Model Family (2026)

Use these as launch points, then tune in 0.5 increments. Always verify against the model card — community fine-tunes vary widely.

Model family	Typical CFG	Notes
SD 1.5 / SDXL derivatives	5–8	Stylized/anime fine-tunes often prefer 6–8; photorealism 3.5–6.
SD3.5-class models	4–7	More sensitive to extremes than SDXL.
Flux-class (distilled guidance)	1.0 CFG + guidance 2.5–4	Raise the guidance parameter, not CFG.
Turbo / LCM / few-step models	1.0–2.0	Guidance is distilled in; high values destroy coherence.

Negative Prompts and CFG Are a Coupled System

CFG amplifies the entire conditioning vector, including your negative prompt. This creates a counterintuitive effect: adding an aggressive negative prompt lets you run a lower CFG while still suppressing unwanted traits. Stacking a heavy negative prompt and a high CFG double-applies the same correction, which is exactly how you get artifacts.

Practical rule: when output looks overcooked, try lowering CFG first, then softening the negative prompt — not the other way around.

Advanced Guidance Techniques Worth Knowing

Guidance rescaling and dynamic thresholding

Techniques like CFG rescaling, zero-star guidance, and dynamic thresholding normalize the guided prediction to prevent pixel value explosion. If your UI exposes these, they let you push CFG slightly higher without clipping.

CFG scheduling and guidance intervals

Guidance rarely needs to be constant across all sampling steps. Applying strong guidance during early steps (composition) and weaker guidance late (texture and detail) produces cleaner results than a single fixed value. Many 2026 node graphs expose a guidance schedule curve for this reason.

Alternative guidance methods

Newer samplers replace raw CFG extrapolation with more stable variants:

• APG (Adaptive Projected Guidance) — projects out the components of the guidance vector that cause oversaturation.
• TCFG (Tangential CFG) — removes the component of guidance correlated with the unconditional prediction, reducing color burn.
• SLG (Skip Layer Guidance) — applies guidance to a subset of network layers, improving anatomy without global contrast damage.
• Perturbation guidance — uses degraded or self-attention-perturbed conditioning instead of a true unconditional pass, cutting inference cost.

Practical Tips for Tuning CFG Fast

• Lock the seed. Change only CFG between renders, or you are comparing noise, not guidance.
• Move in 0.5 steps, and sweep 3–9 in a single XY plot when learning a new checkpoint.
• Fix prompts before touching CFG. A high guidance value cannot rescue vague language — use token weighting or regional prompting instead.
• Mind your sampler. Ancestral samplers (Euler a, DPM++ 2S a) tolerate higher CFG; deterministic DPM++ 2M shows oversaturation earlier.
• Lower CFG for realism. Natural skin, film grain, and soft light live in the 3.5–5.5 band on most photoreal checkpoints.
• Raise CFG for graphic styles — logos, flat vector art, and typography often benefit from firmer adherence.
• Re-tune when you change the model. Guidance behavior is a property of training, not a universal constant.

Common Mistakes to Avoid

• Assuming “higher CFG = better prompt following.” Past the optimum, adherence degrades.
• Carrying SD1.5 habits (CFG 7–9) into distilled models, where those values are catastrophic.
• Compensating for a weak prompt with CFG instead of sharper, more specific language.
• Ignoring the interaction between CFG, sampler, and step count — all three shape the final look.

Conclusion

CFG scale and guidance are not quality dials — they are adherence dials with a narrow optimum that shifts with every model, sampler, and prompt. In 2026, the skill is less about finding a magic number and more about understanding which guidance system you are actually adjusting: classic classifier-free guidance, distilled guidance, or a rescaled projection method like APG or TCFG. Start at the model’s recommended range, change one variable at a time, keep your seed fixed, and let prompt craft do the heavy lifting. Get that balance right, and the guidance slider becomes a precision instrument instead of a wrecking ball.