CFG Scale & Guidance in AI Art: 2026 Guide

You’ve written a clean prompt, hit generate, and the model hands back a blown-out, over-saturated mess — or worse, an image that ignores half your words. Nine times out of ten, the culprit isn’t your prompt. It’s the CFG scale. Classifier-free guidance remains the single most powerful lever for AI image control in 2026, and it’s also the most misunderstood slider in every interface that exposes it. This guide breaks down what guidance scale actually does, how modern architectures handle it differently, and which values to reach for when you need reliable results.

What Is CFG Scale, Exactly?

CFG (classifier-free guidance) is a multiplier applied inside the denoising loop. At every step, the model runs two predictions: one with your prompt attached (the conditional prediction) and one with an empty prompt (the unconditional prediction). The final result is a blend:

output = uncond + CFG × (cond − uncond)

That formula explains everything. At a CFG of 1, you get the raw conditional prediction — your prompt influences the image, but nothing is amplified. At 0, your prompt is effectively ignored. Above 1, the difference between “with prompt” and “without prompt” gets stretched, pushing the image harder toward your text. Below 1, prompt adherence softens and the model’s own aesthetic takes over.

The tradeoff is structural, not incidental. Amplifying a prediction means amplifying its errors too. That’s why high CFG values don’t just mean “more prompt adherence” — they mean more contrast, more saturation, harder edges, and eventually visible artifacts.

CFG Scale vs. Guidance Scale: Same Idea, Different Plumbing

The terms are often used interchangeably, but in 2026 they aren’t always the same parameter. “CFG scale” refers to the classic two-pass classifier-free guidance above. “Guidance scale” is the broader term, and in newer models it may refer to:

• Guidance embedding / distillation — the model was trained to respond to a guidance value as an input (Flux.1 [dev] is the classic example), so true CFG is often left at 1 while a separate guidance field does the work.
• Guidance rescale — a correction factor that compensates for the over-exposure CFG causes, common in SD3.5 and Flux-family workflows.
• Style-based equivalents — Midjourney’s --stylize and --weird control interpretation rather than literal prompt matching, which is a different axis entirely.

What the Guidance Slider Actually Changes

• Prompt adherence — higher values follow your text more literally, including details you didn’t really want.
• Contrast and saturation — rises quickly above the model’s sweet spot; a CFG of 12 looks like an over-sharpened JPEG.
• Texture quality — very high values produce waxy skin, plastic surfaces, halos, and chromatic fringing.
• Variation across seeds — lower values produce more diverse compositions from the same prompt.
• Negative prompt strength — negatives only function when CFG is above 1, and their effect scales with it.

Model-Specific Guidance Ranges in 2026

Defaults vary widely. Treat these as starting points, then check the model card — that’s the source of truth.

• SD 1.5 family: 6–9. Old habits die hard, but anything above 9 is usually damaging.
• SDXL: 4–7. The sweet spot for most work sits around 5–6.
• SD3 / SD3.5: 3.5–6, with CFG rescale enabled. SD3 punishes high CFG harder than SDXL.
• Flux.1 [dev]: guidance 1.5–4 (default ~3.5), true CFG at 1. Flux.1 [schnell] wants guidance at 0.
• Turbo, Lightning, and distilled LoRAs: CFG 1–2 with 4–8 steps. Higher values break these models entirely.
• Modern flow-matching models: typically 3–5, with many now supporting dynamic guidance schedules.
• Midjourney: --stylize 0–1000 (default 100) for artistic interpretation, not prompt literalism.

How CFG Interacts With Your Other Controls

Samplers and step count

Guidance and steps are not independent. Higher CFG converges faster on the prompt but overshoots more easily, so raising guidance while also raising steps tends to produce burnt results. If you push CFG up by 2, consider dropping steps slightly. Conversely, low CFG at low step counts produces soft, undercooked images.

Negative prompts

Long negative prompt lists are a common mistake. At high CFG, a stack of negatives fights the positive prompt and produces warped anatomy, extra limbs, and strange color shifts. Keep negatives to three to six focused terms. If you’re tempted to add a seventh, try lowering CFG by one instead — it usually solves the problem more cleanly.

Distilled and guidance-embedded models

If a model was trained with guidance distillation, traditional CFG is redundant. Setting CFG to 2 or 3 on top of a baked-in guidance value creates double-guidance artifacts. On these models, keep CFG at 1 and tune the dedicated guidance parameter.

Practical Tips for Tuning Your Guidance Scale

• Start at the model default. Before you touch anything else, generate at the recommended value so you have a baseline.
• Change one variable at a time. Guidance, then steps, then prompt. Changing all three makes results unreadable.
• Move in increments of 1, not 5. The difference between 5 and 6 matters more than the difference between 5 and 12.
• If the image looks “cooked,” drop CFG by 1–2 before you rewrite the prompt. Over-guidance mimics a bad prompt.
• If the model ignores a key subject, raise CFG by 1 and move that subject earlier in the prompt — order matters as much as magnitude.
• Enable CFG rescale on SD3 and Flux-family workflows whenever CFG is above 1.
• Try dynamic guidance schedules: higher guidance early for composition, lower guidance late for natural texture.
• For exploration, run CFG 2–3 across multiple seeds. For product renders and text-heavy images, run 5–7 with a fixed seed.
• Never judge a value on one image. Guidance interacts with seeds; test across at least four.

Troubleshooting Common Guidance Problems

• Blown highlights, oversaturation: CFG too high. Reduce by 1–2 or enable rescale.
• Prompt details ignored: CFG too low, or the prompt is ambiguous. Raise by 1 before adding adjectives.
• Extra limbs, mangled hands: high CFG combined with a heavy negative prompt. Cut the negatives first.
• Flat, washed-out images: CFG below 2 on a model that expects 4–6.
• Muddy output on a turbo model: CFG set above 2. Distilled models want