If you have ever typed a prompt into an AI image generator and gotten back something that looked nothing like what you imagined, the CFG scale is often the silent culprit. It is one of the most powerful sliders in your toolkit — and one of the most misunderstood. In 2026, with a dozen different model families each expecting a different guidance range, knowing how to tune this single parameter can mean the difference between a flat, washed-out draft and a crisp, prompt-faithful final render.
This guide breaks down what the CFG scale (classifier-free guidance) actually does, how the guidance scale varies across modern models, and exactly how to use it for precise AI image control.
What Is the CFG Scale?
CFG stands for classifier-free guidance. It is a numerical value that tells the model how strongly it should follow your text prompt relative to its own learned defaults. A low value lets the model improvise. A high value forces it to obey you — sometimes to the point of breaking the image.
The term comes from the original 2022 diffusion research, but the concept has survived into every major architecture in use today: latent diffusion, flow matching, and the distilled turbo models that dominate real-time generation.
How Guidance Works Under the Hood
During each denoising step, the model makes two predictions:
- Conditional prediction — what the image should look like given your prompt
- Unconditional prediction — what the image should look like with no prompt at all
The sampler then extrapolates away from the unconditional guess and toward the conditional one. The CFG scale is the multiplier on that extrapolation. In simplified form:
final = unconditional + CFG × (conditional − unconditional)
At a CFG of 1, you get the raw model output. At 7, you get a strong push toward your prompt. At 20, you get oversaturated, crunchy artifacts as the sampler overshoots the target.
Why It’s Also Called the Guidance Scale
Different ecosystems use different names for the same dial. Stable Diffusion interfaces call it CFG Scale. Flux, SD3.5, and most flow-matching tools call it Guidance Scale. Some front-ends label it “Prompt Strength.” They are functionally equivalent, though the numeric ranges are not interchangeable between model families.
CFG Scale Cheat Sheet by Model Family (2026)
There is no universal “correct” value. Each architecture was trained with a different guidance expectation:
- SD 1.5 / SD 2.x: 7–8 is the classic sweet spot; 5–12 is the usable range
- SDXL Base: 5–8 for most work; 4–6 for photorealistic portraits
- SD3 / SD3.5: 4–7; higher values cause color banding
- Flux.1 dev: 2–4 with distilled guidance; set CFG to 1.0 when distilled guidance is active
- Turbo / LCM / Lightning / DMD2: 1.0–2.0. These are distilled for few-step generation and break at higher values
- Hosted platforms (Midjourney, DALL·E, Firefly): no exposed CFG. They use internal guidance with stylistic controls like stylize or weirdness instead
Practical tip: Before you tune anything else, check the model card. Using a CFG of 7 on a turbo model is the single most common cause of muddy, oversharpened output.
Diagnosing a Bad Guidance Setting
Signs CFG Is Too High
- Blown-out highlights and neon, oversaturated colors
- Hard halos or glowing edges around subjects
- Repeated or fused anatomy — extra fingers, merged limbs
- Texture that looks “fried,” crunchy, or over-sharpened
- Composition feels rigid and literal, with no atmosphere
Signs CFG Is Too Low
- The prompt is ignored or only partially followed
- Output looks flat, hazy, or low-contrast
- Style words like “cinematic” or “watercolor” have little effect
- Negative prompts stop working entirely
If you see symptoms from both columns, your guidance is fine but your step count or sampler is not.
Advanced Guidance Techniques Worth Knowing
By 2026, plain CFG is rarely the only option. Modern pipelines ship smarter guidance methods that solve the classic trade-off between prompt adherence and image quality:
- CFG Rescale: compensates for contrast blowout at high CFG. A rescale value around 0.7 lets you push adherence without frying the image.
- Perturbed Attention Guidance (PAG): applies guidance in the attention layers instead of the noise prediction, improving structure with less color damage.
- Adaptive Projected Guidance (APG): projects the guidance vector to prevent oversaturation — popular in ComfyUI workflows.
- Skip Layer Guidance (SLG): targets specific transformer blocks, useful for anatomy fixes in video and large image models.
- CFG-Zero*: zero-initializes guidance for flow-matching models, eliminating the early-step instability that plagues Flux and SD3 at higher scales.
- Distilled guidance: Flux and similar models bake guidance into training, so you set guidance to a small number and CFG to 1.0.
If you use a node-based interface, these are usually drop-in replacements for standard CFG. Start with one at a time — stacking them compounds unpredictably.
Practical Tips for Tuning Guidance
- Change one variable at a time. Adjust CFG while keeping the seed, sampler, and steps fixed, or you will not know what caused the change.
- Use a grid or XY plot. Generate the same prompt across 3–11 in increments of 2. You will find your model’s real sweet spot in under a minute.
- Write the prompt before you tune. A vague prompt at CFG 7 will beat a great prompt at CFG 3, but a coherent, specific prompt gives guidance something to work with.
- Lower guidance for realism, raise it for illustration. Photoreal skin and fabric look best at 4–6. Graphic styles, logos, and text benefit from 8–10.
- Balance against steps. High CFG with very few steps produces artifacts. If you push above 10, add steps.
- Remember that negatives live in the unconditional branch. Strong negative prompts effectively raise guidance pressure — reduce CFG if negatives are causing distortion.
- Save presets. Once you find working values per model, store them. Your future self will thank you.
How Guidance Interacts With Other Settings
Guidance does not act in isolation. It competes with your sampler choice, step count, scheduler, and prompt complexity. A few rules of thumb:
- DPM++ and Euler-based samplers tolerate higher CFG better than ancestral samplers
- More steps slightly reduce the visible damage of high CFG but never eliminate it
- Very long prompts can behave like a higher guidance scale, because more tokens create stronger conditioning
- LoRA stacks amplify effective guidance — reduce CFG when layering multiple LoRAs
Conclusion
The CFG scale is the steering wheel of AI image generation. It decides how much the model listens to you versus how much it trusts its own training. Get it right and your prompts translate cleanly into images. Get it wrong and you either lose control or lose image quality — there is rarely a middle ground that fixes itself.
For 2026 workflows, the winning approach is simple: learn your model’s native guidance range, start near the low end, and raise the value only until prompt adherence improves without color or anatomy degrading. Pair that with modern techniques like CFG Rescale or PAG when you need extra control, and keep a tested preset for every model you use.
Mastering that one number will improve more of your images than any new model release this year.