If your AI-generated images keep drifting away from your prompt — or come out looking oversaturated, crunchy, and “fried” — the CFG scale is almost always the setting to blame. Classifier-Free Guidance is the single most misunderstood dial in AI image generation, and in 2026 it’s also the most fragmented: some models expose it plainly, others hide it behind embedded guidance, and a growing number of turbo and distilled checkpoints don’t use it at all. Getting this one parameter right can mean the difference between a usable render and a wasted afternoon. This guide breaks down what the CFG scale actually does, how guidance scale differs across modern models, and the practical rules you need for precise AI image control.
What Is the CFG Scale?
CFG stands for Classifier-Free Guidance, a technique introduced by Ho and Salimans and now baked into nearly every diffusion pipeline. When you generate an image, the model effectively makes two predictions at each denoising step:
- A conditional prediction — what the image should look like given your prompt.
- An unconditional prediction — what the image would look like with no prompt at all (or with your negative prompt).
The CFG scale is the multiplier applied to the difference between those two. Push it higher, and the model leans harder into your prompt and away from the generic “average image” prior. That’s why high CFG values improve prompt adherence — and why they also amplify contrast, saturation, and eventually artifacts.
How Guidance Scale Differs From CFG
In 2026, “CFG scale” and “guidance scale” are often treated as synonyms, but they aren’t always the same thing:
- True CFG requires two forward passes per step (conditional and unconditional), which roughly doubles generation time.
- Embedded or distilled guidance — used by Flux.1 [dev] and similar models — trains guidance directly into the network, so a single pass is enough. The guidance value still matters, but its sweet spot is much lower.
- No guidance at all — turbo, LCM, Lightning, and DMD2-distilled models are guidance-distilled and expect a CFG of exactly 1.
This is why copying settings from one model to another produces garbage. A CFG of 7 is normal on SDXL and catastrophic on Flux Schnell.
2026 CFG Scale Cheat Sheet
Use these as starting points, then fine-tune in increments of 0.5 to 1.
| Model family | Recommended CFG / guidance |
|---|---|
| SD 1.5 / SDXL base | 5 – 8 |
| SD 3.5 Large | 3.5 – 5 |
| Flux.1 [dev] | 1.5 – 4 |
| Flux.1 [schnell], turbo, LCM, Lightning | 1 |
| Midjourney / most hosted APIs | Not exposed — tune stylize or equivalent instead |
As a general rule: the more modern and more distilled the model, the lower the guidance scale it wants.
What Different CFG Values Actually Look Like
Low CFG (1 – 4)
- Soft, painterly, dreamlike results with loose prompt adherence.
- Greater variety between seeds — useful for ideation and concept exploration.
- At CFG 1 on a standard model, the negative prompt stops working entirely and output turns muddy.
Mid CFG (5 – 8)
- The classic balance zone for older architectures: recognisable subjects, natural saturation, stable anatomy.
- Best default for portraits, environments, and general illustration.
High CFG (10 – 20+)
- Aggressive prompt adherence — helpful for precise compositions, product mockups, and text rendering.
- Side effects: blown highlights, neon saturation, haloed edges, duplicated limbs, and reduced image diversity.
- Beyond roughly 15 on most models, quality collapses regardless of step count.
How CFG Interacts With Your Other Settings
Sampler and Step Count
High CFG overshoots the target and can converge faster, so some workflows pair CFG 9–12 with fewer steps. In practice, lowering CFG is almost always the better fix. Modern samplers like DPM++ 3M SDE and UniPC handle mid-range CFG far more gracefully than the old Euler a.
Negative Prompts
Negative prompts are pure CFG mechanics: they only exist when CFG is above 1. If you’re running a distilled model at CFG 1, delete the negative prompt node — it’s dead weight. If you’re at CFG 7+, a focused negative prompt (“blurry, extra fingers, watermark”) works far better than a 40-token monster list that fights your positive prompt.
CFG Rescale and Dynamic Guidance
Two 2026 staples worth enabling:
- CFG Rescale — a post-step correction (typically 0.7–0.9) that counteracts the contrast blowout at high CFG, letting you push adherence without the “fried” look.
- Dynamic or scheduled guidance — runs high guidance early to lock composition, then drops it mid-generation to restore natural texture and detail.
Practical Tips for Better AI Image Control
- Change one thing at a time. Lock your seed, then sweep CFG values (3, 5, 7, 9, 12) and compare side by side.
- Think of CFG as a fidelity dial, not a quality dial. More is not better — it’s just more literal.
- Match the model. Check the model card before assuming a default. Distilled checkpoints at CFG 7 will look burnt.
- Compensate, don’t compensate twice. If you raise CFG for prompt accuracy, consider lowering contrast prompts or enabling rescale rather than adding more steps.
- Go low for creativity, high for precision. Brainstorming at CFG 3 beats grinding at CFG 12.
- Save presets. Store a CFG/sampler/steps trio per model in your UI so you stop re-tuning from scratch.
Common CFG Mistakes to Avoid
- Cranking CFG to 15+ to “force” a difficult prompt instead of rewriting it.
- Leaving CFG at 7 while switching to a turbo or Flux Schnell checkpoint.
- Blaming the sampler for artifacts that are really a guidance problem.
- Expecting negative prompts to work in hosted tools that don’t expose guidance at all.
Conclusion
The CFG scale remains the most powerful lever for AI image control in 2026 — but its era of one-size-fits-all defaults is over. Modern architectures span true classifier-free guidance, embedded guidance, and fully distilled models that want CFG 1, so the real skill is knowing which regime you’re in. Start at your model’s recommended range, sweep in small increments with a fixed seed, and reach for CFG rescale or scheduled guidance before you reach for extreme values. Treat guidance scale as a fidelity dial rather than a quality slider, and you’ll get images that follow your prompt without the telltale fried-pixel tax.