Open any Stable Diffusion interface — ComfyUI, Automatic1111, InvokeAI, the HyperFlow-style distilled models covered elsewhere on this site — and two controls sit right next to your prompt box: CFG scale, usually a number between 1 and 20, and a negative prompt field. Almost everyone touches both immediately. Almost nobody is taught what either one actually does, which is why “just turn CFG up for more prompt adherence” is both roughly true and a great way to ruin an image.
What CFG scale actually controls
Classifier-free guidance (CFG) is not a filter or a strength slider in the way a Photoshop opacity value is. At every denoising step, a CFG-enabled model runs the network twice: once conditioned on your prompt, and once unconditioned (or conditioned on an empty/negative prompt). The final prediction is the unconditional prediction pushed in the direction of the conditional one, scaled by the CFG value:
final = unconditional + CFG_scale × (conditional − unconditional)
At CFG 1, the model essentially ignores your prompt and generates something plausible-but-generic. As CFG climbs, the model is pushed harder away from “generic plausible image” and toward “an image that scores well against your specific prompt.” That sounds like it should mean “higher CFG = better prompt following, always” — and up to a point, it does. But because CFG is extrapolating a vector, not applying a clean filter, pushing it too hard overshoots: colors blow out and saturate unnaturally, contrast gets harsh, and fine structure (hands, text, fabric texture) starts to warp because the extrapolation is amplifying noise along with signal. This is why most models have a practical sweet spot — commonly cited as somewhere around 3-8 for modern models, far lower than the 7-12 range that was standard advice for older Stable Diffusion 1.x checkpoints — and why newer distilled and few-step models increasingly run at CFG 1 by design, baking prompt adherence into the model itself rather than achieving it through guidance at inference time.
What a negative prompt actually subtracts
A negative prompt does not work by telling the model “don’t generate this.” The model has no mechanism to understand negation as a concept; “no clouds” and “clouds” both put the token “clouds” into its attention.
What actually happens: your negative prompt text replaces the unconditional branch in the CFG equation above. Instead of guidance pointing away from “a generic plausible image,” it points away from “an image that matches this negative description.” So a negative prompt of blurry, extra fingers, watermark doesn’t erase blur, extra fingers, or watermarks from existence — it steers the final prediction further from whatever region of the model’s learned space produces those things, which is a much weaker and more probabilistic effect than people expect. It’s a nudge, not a mask.
This is also why stuffing a negative prompt with generic quality terms (bad quality, worst quality, low resolution) has diminishing and sometimes counterproductive returns: each additional negative term dilutes the unconditional branch’s focus, and at high CFG values, a long, unfocused negative prompt can introduce its own instability rather than removing one.
Practical guidance
- Raise CFG when your generations ignore your prompt, not as a default “make it better” move. If increasing CFG by 1-2 points doesn’t visibly improve prompt matching, the fix is probably in the prompt itself, not the slider.
- Lower CFG when you see blown-out colors, harsh contrast, or warped fine detail at a given CFG value — these are classic guidance-overshoot symptoms, and the fix is often just turning the number down 2-3 points before changing anything else.
- Use negative prompts for things the model demonstrably tends toward, not an ever-growing junk list. A targeted negative prompt (specific artifacts you’re actually seeing in your outputs) does more than a generic “ugly, deformed, bad anatomy” block copied from a tutorial.
- Check whether your model wants CFG at all. Several newer, few-step and distilled video and image models are explicitly designed to run with CFG disabled (effectively CFG 1) — applying old-model guidance advice to them will make results worse, not better, because the model was never trained expecting that extrapolation.
Neither control is a quality dial. Both are steering mechanisms on top of a model that already has its own strong priors about what a plausible image looks like — which is exactly why understanding the mechanism, rather than copying someone else’s numbers, is what actually transfers across models.