The right guidance setting in each tool, and how to tell when it's wrong

Know what you're turning

Classifier-free guidance (CFG) is the number that sets how hard the model follows your prompt. At each step the model predicts once with your prompt and once without it (or with your negative prompt), then moves further in the direction your prompt pulls. Higher number, harder push.

Too high and the colours burn: Google's Imagen paper describes high guidance weights giving "highly saturated and unnatural images". Too low and the picture drifts into something generic that half-ignores your words.

Not every tool has the dial. Some models are distilled so the guidance is baked in, and the big chat-style tools hide it completely.

Find the setting in your tool

Automatic1111 and Forge: the CFG Scale slider under the prompt (1 to 30, default 7). ComfyUI: cfg on the KSampler node (default 8.0).

Flux on fal: guidance_scale (default 3.5). On Replicate's flux-dev: guidance (default 3.5, maximum 10). BFL's own FLUX.2 [flex] endpoint: guidance (1.5 to 10, default 5). Stability AI's SD 3.5 API: cfg_scale (1 to 10).

No dial at all: Midjourney, ChatGPT images (the OpenAI Image API has no guidance parameter), Nano Banana, Ideogram, FLUX.2 [pro] and Kling's current 3.0 video API. Skip to the step on tools without a dial.

Start from the right number for your model

Stable Diffusion 1.5: 7 (Automatic1111's default; diffusers' default is 7.5). SDXL: start between 5 (the diffusers default for SDXL) and 7. SD 3.5 Large and Medium: 4 on Stability's API (its default), 3.5 to 4.5 in Hugging Face's own examples. SD 3.5 Turbo and Flash: 1.

Flux dev: guidance 3.5. Replicate's notes for flux-dev say "lower values can give more realistic images" and suggest trying 2, 2.5, 3 and 3.5. FLUX.2 [flex]: 5 is BFL's default.

Flux schnell is distilled to work in 1 to 4 steps with guidance at 0, so there's nothing to tune there.

Flux in Forge and ComfyUI: two dials, set one

Flux dev is guidance-distilled: the steering is trained into the model and you set it with a separate number. Forge's own Flux announcement says to set CFG Scale to 1 and use Distilled CFG Scale instead, default 3.5.

In ComfyUI, set KSampler cfg to 1.0 (ComfyUI's Flux examples say so) and set guidance on the FluxGuidance node, which also defaults to 3.5.

At CFG 1 the second guess is skipped. Forge's log says it outright: "Negative Prompts are ignored." So with Flux dev, put what you don't want into the positive prompt as what you do want ("clean empty background" rather than a negative "clutter").

Run a five-image sweep

Guess less, test more. Lock the seed so the only thing changing is guidance, then run five values around the starting point. For SDXL: 3, 5, 7, 9, 12. For Flux dev: 2, 2.5, 3.5, 5, 7.

Put them in a row. Somewhere in the middle the picture follows your prompt without looking cooked. That's your number for this kind of prompt.

I want to find the right guidance value. Tool: [Forge / ComfyUI / fal / Replicate / BFL API]. Model: [e.g. SDXL, Flux dev, FLUX.2 flex, SD 3.5 Large]. Prompt: [PASTE]. Tell me the setting's exact name in this tool, a sensible starting value with where that number comes from, and five values to sweep on a locked seed. Then give me a small table to note for each: follows the prompt (yes/partly/no), colour (natural/strong/burnt), detail (soft/good/harsh).

Read the result

Too high: oversaturated colour, harsh contrast, blown or crushed tones, an unnatural, overcooked look. Lower it by 1 to 2 (or by 0.5 for Flux).

Too low: muddy or washed-out colour, missing objects from the prompt, the subject drifts. Raise it by the same steps.

Still ignoring a word at a high value? That's a prompt problem, not a guidance problem. Move the word earlier, describe it more concretely, or remove words that fight it.

In tools without a dial, steer with the prompt

Midjourney: --stylize (or --s) runs from 0 to 1000, default 100. Midjourney describes low values as following the prompt more closely and high values as more artistic and free. For more literal results, lower --s or turn on Raw, which removes default styling.

ChatGPT images and Nano Banana: there's no dial, so the prompt is the only lever. Be specific: say what must be there, where, and what must not change. If an image ignores something, ask for that one fix in the same chat rather than rewriting everything.

Ideogram: no guidance control either. Its style type (in the API: Auto, General, Realistic, Design or Fiction) and the negative prompt change what it aims for, and Ideogram notes the prompt takes precedence over the negative prompt.

“A coffee shop.”

A specific brief

Why a negative prompt sometimes does nothing

In Stable Diffusion-style tools the negative prompt replaces the "without your prompt" guess. Guidance then pushes away from it. That's why it stops working at CFG 1, the setting for Flux dev and for turbo models: there's no second guess to push away from.

Save your numbers

Keep a note of the value that worked per model and per kind of image (portrait, product, illustration). Photoreal work usually wants the lower end of a model's range, graphic and illustrated styles can take more. Check that against your own sweep, not a forum post.

Turn my sweep notes into a one-page settings card for my tools: model, setting name, my best value per image type (portrait, product, illustration, text-heavy), and the signs that tell me to go up or down. Notes: [PASTE].