CFG Scale & Guidance in AI Image Generation 2026
The single slider that changes more AI images than any other is the CFG scale. Set it too low and your prompt is politely ignored; set it too high and you get fried colors, crunchy textures, and the AI equivalent of someone shouting every word of a sentence. Yet in 2026, the guidance scale behaves very differently depending on which model you load — Flux, SDXL, SD3.5, and the new distilled turbo checkpoints all want completely different values. This guide explains what the CFG scale actually does, what sensible values look like per model, and how to use guidance as a precise tool for AI image control instead of a mystery number you copy from a Reddit thread.
What Is the CFG Scale?
CFG stands for classifier-free guidance. It’s a parameter — usually ranging from 1 to 20 — that controls how strongly your text prompt steers the diffusion process. Internally, the model makes two predictions at every denoising step:
- The conditional prediction: what the image should look like given your prompt.
- The unconditional prediction: what the image should look like with no prompt at all (or with your negative prompt instead).
The CFG scale determines how far the final output is pushed away from the unconditional version and toward the prompt-driven one. Low values keep things dreamy, varied, and loosely related to your prompt. High values force a tighter match — at the cost of image quality.
You’ll see this setting labeled guidance scale, text guidance, or prompt strength depending on the interface. They mean the same thing, with one important exception covered below.
How Guidance Scale Actually Works
The Math in Plain English
The core formula looks roughly like this: final prediction = unconditional + scale × (conditional − unconditional). When the scale is 1, the model uses only the prompt-conditioned prediction. When the scale is 7, the difference between “prompt” and “no prompt” gets multiplied by seven before being applied. That amplified difference is why high CFG produces contrasty, saturated, aggressively literal images.
Why Higher Isn’t Better
Guidance amplifies the difference between two noisy predictions — and noise gets amplified too. Push the scale past a model’s comfortable range and you’ll see predictable damage:
- Blown-out highlights, neon oversaturation, and crushed shadows
- Hard halos or “cutout” edges around subjects
- Plastic skin, distorted anatomy, and repetitive textures
- Reduced variety across a batch — every seed converges on the same literal interpretation
The sweet spot is always the lowest value that makes the model obey your prompt. Not the highest value you can survive.
Typical CFG Ranges by Model (2026)
This is where most people go wrong: guidance values are not portable between architectures.
- SD 1.5 and SD 2.x: 7–12. The classic default of 7 still works, but these older models tolerate 10–12 better than newer ones.
- SDXL and its derivatives: 5–8. Around 6–7 is the usual sweet spot; 8+ starts to break down skin and fine detail.
- SD3 / SD3.5: 4–6.5. These models were trained with lower guidance and become brittle above 7.
- Flux.1 dev: 2.5–4 (default is typically 3.5). Flux has a built-in guidance embedding, so the external slider is far less aggressive.
- Flux schnell, SDXL Turbo, Lightning, and other distilled models: 1–2, or none at all. These are guidance-distilled — external CFG often does nothing or actively harms output.
- Midjourney: no CFG slider. Use
--stylize(defaults around 100) for prompt adherence versus artistic license, and--chaosfor variety. - DALL·E 3 / GPT-image class models: guidance is handled internally. Instead of a slider, control adherence with prompt specificity and instruction phrasing.
How CFG Interacts With Your Other Settings
Steps and Samplers
Higher CFG needs more steps to resolve cleanly, because each step applies a stronger correction. If you raise CFG, raise steps slightly. Modern samplers like DPM++ 2M, Euler a, and the current 2026 generation of flow-matching schedulers all have slightly different tolerance — a sampler that looks great at CFG 6 may look harsh at CFG 9.
Negative Prompts
The negative prompt is the unconditional prediction at higher CFG values. This means negative prompts only do meaningful work when CFG is above roughly 3–4. On Flux and other low-guidance models, negatives are often ignored entirely.
LoRAs and Fine-Tunes
LoRAs trained at CFG 7 tend to want CFG 7 at inference. Community checkpoints are usually tuned to the base model’s default, so start at default and adjust from there. If a LoRA’s effect seems weak, lower the CFG rather than increasing the LoRA weight — high CFG fights subtle stylistic training.
Practical Tips for Better AI Image Control
- Change one variable at a time. Lock your seed, then sweep CFG in steps of 1 to see the real effect.
- Start low, climb slowly. Begin at the model’s low end. If the prompt is being ignored, add 0.5, not 3.
- Fix prompts before you fix guidance. “Ignored prompt” is usually a vague prompt, not a CFG problem. Be concrete about subject, lighting, lens, and composition.
- Use dynamic CFG if your UI supports it. Scheduling a higher scale early and tapering it later gives strong composition with cleaner detail.
- Try CFG++ or guidance rescaling. Both techniques reduce the quality penalty of high guidance and can let you run 2–4 where you used to need 7.
- Batch compare. Generate four seeds at CFG 4, 6, and 8 at identical settings. Patterns appear fast.
- Remember ControlNet. ControlNet weight is a second guidance system. If you’re already using strong structural guidance, you usually need less CFG.
Common Problems and Quick Fixes
| Symptom | Likely Cause | Fix |
|---|---|---|
| Washed-out, flat images | CFG too low | Increase by 1–2 |
| Oversaturated, crunchy detail | CFG too high | Decrease by 1–3 |
| Prompt elements missing | Low CFG or vague wording | Raise slightly, then rewrite the prompt |
| Negative prompt ignored | CFG below ~3 | Raise to 5+ or drop negatives |
| Identical, boring batches | CFG too high | Lower to 4–5 and vary seeds |
The 2026 Shift: Guidance Is Getting Smarter
The trend across 2025–2026 is clear — manual CFG tuning is fading for casual users. Distilled models bake guidance into training, and interfaces increasingly offer automatic guidance that adapts per prompt and per step. Per-region guidance, where you apply different strengths to different parts of the canvas, is now practical in several editors. For professional work, though, understanding the CFG scale remains the difference between a model that obeys you and one that merely approximates what you asked for.
Conclusion
The CFG scale is not a quality dial — it’s a compliance dial. Treat it that way and your results improve immediately. Use the lowest value your model respects, learn the per-architecture ranges (7–12 for SD 1.5, 5–8 for SDXL, 4–6.5 for SD3.5, 2.5–4 for Flux dev, and 1–2 for distilled models), and remember that negative prompts, samplers, and ControlNet all interact with guidance. Master that balance and you gain genuine, repeatable AI image control — no guesswork required.