Ask ten AI artists which setting they reach for first when an image refuses to cooperate, and most will give the same answer: the CFG scale. Also known as the guidance scale, this single number decides how aggressively a diffusion model follows your prompt — and it shapes everything from color accuracy and contrast to whether a result looks natural or “deep-fried.” Yet in 2026, with guidance-distilled architectures, new schedulers, and cloud platforms that hide the dial entirely, the old advice (“just use 7”) no longer holds. This guide explains what the CFG scale actually does, how it behaves across modern models, and how to tune it for reliable, repeatable AI image control.
What Is the CFG Scale?
CFG stands for classifier-free guidance. During sampling, a diffusion model predicts noise twice: once with your prompt attached (the conditional prediction) and once without it (the unconditional prediction). The CFG scale controls how far the final result leans toward the prompt-driven version. The simplified formula looks like this:
final = unconditional + CFG × (conditional − unconditional)
At a CFG of 1.0, the guidance term is nearly neutral, so the model wanders freely. As the value climbs, the prompt direction gets amplified. Push it too far and the math overshoots, which is why extreme values produce the tell-tale artifacts every experienced user recognizes.
How Guidance Actually Changes an Image
- Low CFG (1–4): Softer, dreamier, more varied outputs. Great for abstract or painterly work, but the model may quietly ignore specific details in your prompt.
- Moderate CFG (4–8): The balanced zone. Detail, lighting, and composition stay coherent without unnatural contrast.
- High CFG (9+): Literal prompt adherence at the cost of saturation, halos, crunchy texture, and distorted anatomy.
The CFG Scale Sweet Spot: What the Numbers Mean
There is no universal “correct” value, because the scale is relative to each model and sampler. As a starting reference for 2026:
- 2.5–4.5: Flux-family and other guidance-distilled models.
- 4.0–7.0: SDXL, SD 3.5, and most modern photoreal checkpoints.
- 7.0–8.0: Legacy SD 1.5 models, which were trained with this range in mind.
- 9.0–12.0: Only useful when a specific prompt element keeps getting dropped — and usually a sign the prompt itself needs rewriting.
Community finetunes shift these numbers. A stylized anime checkpoint may shine at 5, while a heavily merged photorealism model might peak at 3.5. Treat every published range as a starting point, not a rule.
How Guidance Behaves Across 2026 Models
Flux, SD 3.5, and Guidance-Distilled Diffusion
Modern architectures often bake guidance directly into training. In Flux.1 dev and its successors, you adjust a dedicated guidance parameter rather than a true CFG scale, and the effective range is far lower — typically 1.5 to 5, with 3.5 as the default. Setting it to 7 doesn’t make the model “try harder”; it flattens contrast and degrades fine detail. Some SD 3.5 variants behave similarly when a distilled guidance variant is loaded.
Midjourney, DALL·E, and Hosted Platforms
Cloud tools hide the raw dial but expose equivalents. Midjourney’s --stylize plays a similar role: low values push closer to your literal prompt, high values give the model more aesthetic freedom. --weird and --chaos introduce variation rather than adherence. Understanding that mapping lets you translate local-workflow instincts into hosted tools.
Local Interfaces: ComfyUI, Automatic1111, Forge, InvokeAI
In node-based tools, guidance usually appears in two places: the sampler node’s CFG input and a model-specific node such as FluxGuidance. Feeding a Flux model through a standard CFG input set to 7 is one of the most common mistakes new ComfyUI users make.
Practical Tips for Better AI Image Control
- Start at the model default and change one variable at a time. Adjusting CFG, steps, and sampler simultaneously tells you nothing about which change worked.
- Lock your seed when comparing values. Otherwise you’re comparing random noise, not guidance strength.
- Fix the prompt before raising the dial. Specific, well-ordered prompts outperform brute-force guidance almost every time.
- Lower CFG first when you see artifacts. Halos, oversaturation, and cracked skin texture are guidance problems, not sampler problems.
- Balance steps and CFG. High guidance with 15 steps produces mush. Push steps to 25–35 before going above 8.
- Use negative prompts strategically. A targeted negative (“extra fingers, watermark”) beats a blunt CFG increase.
- Reduce CFG for img2img and inpainting. These workflows already inherit structure from the source image; 3–5 is often plenty.
- Save presets per model. Keep a note of the CFG, sampler, and step count that works for each checkpoint — it will save hours over a year of generation.
Troubleshooting Common CFG Problems
- Prompt ignored, colors washed out: CFG is too low, or the prompt is too vague to guide the model.
- Harsh contrast and glowing edges: CFG is too high — drop by 1.0 and re-test.
- Duplicated limbs or fused faces: Often high CFG combined with weak anatomical cues. Lower guidance and add negatives.
- Style collapse in img2img: Lower the CFG and rebalance the denoise strength.
Advanced Techniques: Scheduling, Rescaling, and Guidance Tricks
Once the basics are second nature, a few advanced methods extend what guidance can do:
- CFG scheduling: Start high (8–10) for the first 30% of steps to lock in composition, then decay to 3–4 for texture and realism.
- CFG rescale: A post-step multiplier (typically 0.5–0.7) that counteracts the contrast burn of high guidance while preserving adherence.
- Dynamic thresholding: Mimics high-CFG prompt adherence on SD 1.5 models without the usual artifacts.
- Regional guidance: Pair ControlNet or regional prompting with per-region values so one element can be pushed hard while the rest stays soft.
- CFG++ and perturbation variants: Newer samplers that reformulate guidance to reduce overshoot, making values above 8 genuinely usable.
Conclusion
The CFG scale remains the most direct lever you have over how faithfully a model interprets your intent — but it is a lever, not a magic wand. The 2026 landscape rewards a model-aware approach: use low guidance values for distilled architectures like Flux, stay in the 4–7 band for SDXL and SD 3.5, and treat anything above 9 as a last resort. Combine that with locked seeds, balanced step counts, thoughtful negative prompts, and scheduling when you need fine control, and your hit rate will climb dramatically. Master the guidance scale, and every other aspect of AI image control — samplers, prompts, and LoRAs — becomes easier to reason about.