How to Use CFG Scale & Guidance in AI Art 2026
If there is one setting that separates a muddy, oversaturated mess from a crisp, prompt-faithful image, it is the CFG scale. Yet in 2026, the rules have shifted. Guidance-distilled models, embedded guidance values, and a new generation of turbo and lightning checkpoints mean the “just use 7” advice you read two years ago will actively hurt your output. This guide breaks down what the CFG scale and guidance scale actually do, how they differ across modern models, and the exact ranges and workflows you should use to get precise AI image control without sacrificing creativity.
What Is CFG Scale (Classifier-Free Guidance)?
CFG stands for Classifier-Free Guidance. During sampling, the model predicts two things at every step: what your image looks like with your prompt, and what it looks like without it (the unconditional prediction). The CFG scale is the multiplier applied to the difference between those two predictions.
In plain terms: the CFG scale controls how aggressively the model pushes the image toward your prompt and away from the generic average of its training data.
- Low CFG (1–3): the image drifts. You get more variety, softer prompt adherence, and often a dreamier, more painterly look.
- Mid CFG (4–8): the sweet spot for most checkpoints — good prompt adherence while keeping natural color and texture.
- High CFG (10+): the model over-commits. Colors turn neon, contrast blows out, and you start seeing “burned” artifacts, halos, and repetitive compositions.
The trade-off you are really making
Think of the CFG scale as a leash. A short leash keeps the model obedient to your prompt but limits how far it can explore a beautiful, unexpected solution. A long leash lets it roam — sometimes brilliantly, sometimes into nonsense. Most of your CFG tuning is about finding the shortest leash that still lets the model make good art.
Guidance Scale vs. CFG Scale: What Changed in 2026
These terms were once interchangeable, but modern architectures have split them apart. The distinction matters:
- CFG scale is applied at inference time by you, using the negative/unconditional branch. It is a slider in ComfyUI, Automatic1111, Forge, InvokeAI, and similar tools.
- Embedded guidance is baked into the model during training. Flux.1, for example, takes a “guidance” parameter that is part of the model’s own conditioning rather than a classic dual-pass CFG.
- Guidance-distilled models (turbo, lightning, LCM, and most 2026 speed checkpoints) were trained to work with no or very low guidance. Pushing CFG to 7 on these models produces fried images instantly.
Meanwhile, platforms like Midjourney have eliminated the CFG slider entirely, replacing it with style and weirdness parameters. If you are on a hosted platform, “guidance” may simply not be exposed — in which case, control comes from prompt weighting instead.
Recommended Guidance Ranges by Model Family (2026)
| Model type | Starting CFG / guidance | Notes |
|---|---|---|
| SD 1.5 & SDXL base | 5–9 (SDXL: 4–7) | Classic CFG. SDXL oversaturates faster, so stay lower. |
| SD 3.5 / newer rectified-flow models | 3.5–7 | Responds more gently than SDXL — test 4.5 and 6.5. |
| Flux.1 (dev / schnell) | 2.5–4 | Uses embedded guidance; above 5 gets plasticky. |
| Turbo / Lightning / LCM checkpoints | 1–2 | Often best at exactly 1.0. Higher values break them. |
| Midjourney and hosted UIs | N/A | Control via style/chaos weighting and prompt specificity. |
Treat these as launch points, not laws. Every fine-tune, LoRA stack, and sampler combination shifts the ideal value by a point or two.
Practical Tips for Dialing In Guidance
1. Change one variable at a time
Lock your seed, prompt, sampler, and step count. Then run an XY plot with CFG on one axis. You will see the exact point where contrast starts to spike — that is your ceiling.
2. Match CFG to your step count
Higher CFG pushes harder per step, so it can compensate for fewer steps — but at the cost of artifacts. A good workflow: raise CFG when stepping down to 20–25 steps, lower it when you have 40+ steps to spare.
3. Treat negative prompts as a CFG amplifier
Long, aggressive negative prompts behave like a bump to your CFG scale. If you add ten negatives and the image gets crunchy, lower CFG rather than trimming the list.
4. Lower CFG when using multiple LoRAs
LoRAs already nudge the model hard. Stacking three or four at high CFG is the fastest route to a burnt, over-detailed image. Drop to the lower end of your range and let the LoRAs do the work.
5. Use guidance for concept exploration, then tighten
Start at CFG 3–4 to explore composition and mood, pick your seed, then raise guidance to 6–7 to lock in prompt fidelity. This two-phase approach beats guessing at a single value.
6. Watch for the four burnout symptoms
- Neon or clipped colors in skin tones and skies
- A hard, embossed outline around subjects
- Repeating patterns or duplicated limbs
- Loss of fine texture — everything looks like plastic
If you see any of these, drop guidance by 1–2 points before changing anything else.
Advanced Workflows: Dynamic and Scheduled Guidance
Manual sliders are only the beginning. In 2026, node-based tools let you schedule guidance across the sampling timeline:
- High-early, low-late: strong guidance for the first 30% of steps to establish composition, then ease off so the model can render natural detail.
- Per-region guidance: combine with regional prompting so a detailed character gets higher guidance than a soft background.
- Guidance rescheduling nodes: normalize guidance across samplers so switching from Euler to DPM++ does not require re-tuning.
These techniques deliver noticeably better AI image control than any single static value, especially in complex, multi-subject scenes.
Common Mistakes to Avoid
- Using CFG 7 on a distilled model. This is the single most common error in 2026 workflows.
- Believing higher CFG equals “better prompt following.” Past a point, it distorts rather than clarifies.
- Ignoring the sampler. Some samplers are far more sensitive to guidance than others — always re-test after switching.
- Never testing the extremes. Running one image at CFG 1.5 and one at 12 teaches you more about your model than fifty images at 7.
Conclusion
The CFG scale and guidance scale remain the most powerful levers for AI image control — but they are no longer one-size-fits-all. In 2026, the correct value depends entirely on your model family: expect 1–2 for turbo and lightning checkpoints, 2.5–4 for Flux, 4–7 for SDXL, and 5–9 for legacy SD 1.5 pipelines. Start at the low end of your range, run a fixed-seed XY plot, and raise guidance only until your prompt is honored without the colors burning out. Master that discipline, add scheduled guidance for advanced work, and you will spend far less time re-rolling and far more time making images that actually match the picture in your head.