CFG Scale & Guidance in AI Image Generation (2026 Guide)
If you have ever typed a perfectly detailed prompt into Stable Diffusion, Flux, or an SDXL workflow and received a washed-out, generic, or strangely over-saturated image, the culprit was almost certainly the CFG scale. Classifier-free guidance is the single most powerful dial in text-to-image generation — and the one most often left at a default value nobody can explain. This 2026 guide breaks down what the CFG scale actually does, how modern guidance scale settings differ across model families, and how to tune it for precise AI image control.
What Is CFG Scale in AI Image Generation?
CFG stands for Classifier-Free Guidance. It is a numerical value that tells the diffusion model how strongly it should follow your text prompt versus how freely it should improvise.
A Quick Definition
During sampling, a diffusion model makes two predictions at every step: one conditioned on your prompt, and one unconditional (the “null” prediction). The CFG scale determines how aggressively the model pushes the final result toward the prompted direction and away from the unconditional one. A low value means loose interpretation; a high value means strict, literal adherence.
How the Math Shapes Your Image
The operation is deceptively simple: the final prediction equals the unconditional prediction plus the CFG scale multiplied by the difference between the conditional and unconditional predictions. Because that difference is amplified, every increment of CFG scale increases contrast between prompt and noise — which is why pushing it too far produces crushed shadows, neon saturation, and melted anatomy.
CFG Scale vs. Guidance Scale: Same Concept, Different Names
Different platforms label the same parameter differently, which causes constant confusion:
- CFG scale — Automatic1111, Forge, ComfyUI, most Stable Diffusion interfaces.
- Guidance scale — Flux, Stable Diffusion 3.5, most APIs (Replicate, Fal, Together), and Hugging Face Diffusers.
- Guidance / CFG in distilled models — often fixed at 1.0 and not adjustable at all.
- Stylize — Midjourney’s proxy for prompt adherence, though its scale is inverted and internal.
If someone says “guidance scale 3.5” and someone else says “CFG 3.5,” they mean the identical parameter.
What Different CFG Values Actually Do
- 1.0 — No guidance. Pure model prior. Required for Turbo, Lightning, LCM, and Flux schnell variants.
- 2.0–4.0 — Soft, painterly, high-variety output. Popular for Flux dev and cinematic renders.
- 4.0–7.0 — The balanced zone for most modern flow-matching and SDXL models. Good prompt adherence with natural color.
- 7.0–9.0 — Legacy Stable Diffusion 1.5 sweet spot. Sharper prompt following, slightly flatter lighting.
- 10.0–15.0 — Increasingly rigid, contrasty, and prone to artifacts. Useful only for specific graphic or poster looks.
- 15.0+ — Near-unusable in most pipelines. Blown highlights, plastic skin, repeated motifs.
Recommended CFG Scale by Model in 2026
| Model family | Typical guidance scale |
|---|---|
| Flux.1 [dev] | 3.0 – 4.0 |
| Flux.1 [schnell] | 1.0 (fixed) |
| Stable Diffusion 3.5 | 3.5 – 5.0 |
| SDXL / SDXL Turbo | 5.0 – 8.0 / 1.0 |
| Stable Diffusion 1.5 | 7.0 – 8.5 |
| LCM & Lightning checkpoints | 1.0 – 2.0 |
Treat these as starting points, not laws. Community checkpoints on Civitai routinely shift the ideal value by one or two points in either direction.
Advanced AI Image Control Techniques
Negative Prompts and CFG
Negative prompts only work through the guidance mechanism — they widen the gap between conditional and unconditional predictions. This means an aggressive negative prompt at high CFG double-amplifies: you get both rigid prompt adherence and rigid negative enforcement, which frequently produces unnatural, over-corrected images. If you use a long negative prompt, lower your CFG scale by one or two points.
Guidance Rescale
Rescale is a post-processing fix that renormalizes the conditional prediction after guidance is applied. It lets you run CFG values of 9 or higher without the characteristic contrast blowout. In ComfyUI, the RescaleCFG node is the standard implementation; most pipelines suggest a rescale factor around 0.7.
Dynamic and Scheduled CFG
Instead of one static value, dynamic CFG schedules guidance strength across sampling steps — typically starting low for composition and rising for detail. Combined with dynamic thresholding, this pushes high-CFG artifacts further down the noise schedule where they matter less.
Distilled and Embedded Guidance
Turbo, Lightning, LCM, and Hyper models bake guidance directly into their training. Setting CFG above 1.0 on these produces severe burning and color collapse. Likewise, recent flow-matching research such as CFG++ shows that reformulating guidance for rectified-flow models improves results at low values — which is exactly why Flux and SD3.5 prefer 3.5 rather than 7.
Practical Tips for Tuning CFG Scale
- Lock your seed. Change only the CFG value across a grid so you can see its isolated effect.
- Use a grid or XYZ plot. Testing 2, 4, 6, 8, and 10 in one pass beats guessing one at a time.
- Pair CFG with step count. Higher guidance needs slightly more steps to resolve cleanly; 25–35 steps is a reasonable range for SDXL and SD3.5.
- Prefer prompt engineering over brute force. If a subject is missing at CFG 5, rewriting the prompt usually beats pushing to CFG 12.
- Match the sampler. DPM++ 2M Karras tolerates higher CFG than Euler a, which tends to burn earlier.
- Check the model card. Almost every major release now publishes a recommended guidance range — it is the fastest shortcut to a good baseline.
Common CFG Mistakes to Avoid
The most frequent error is treating a high CFG scale as a quality setting. It is not. Guidance trades diversity and naturalism for obedience, and past a certain point every additional unit makes images worse, not better. The second most common mistake is ignoring model architecture: applying a Stable Diffusion 1.5 value of 7.5 to a distilled four-step model will destroy the image in a single sampling pass.
Conclusion
CFG scale and guidance scale control the fundamental trade-off at the heart of AI image generation: how literally the model obeys you versus how much natural image quality it preserves. In 2026, the correct answer is almost always lower than beginners expect — 3.5 to 5 for modern flow-matching models, 1.0 for distilled models, and 7 to 8.5 only for legacy SD 1.5 workflows. Master this single parameter, pair it with the right sampler and step count, and your AI image control will improve more than any prompt template ever could.