Skip to content

Learn

What Are Seed and CFG Scale? How They Work

In AI image and video generation, the seed is the number that sets the starting random noise, so the same seed, prompt, model and settings reproduce the same result. The CFG (guidance) scale sets how strictly the model follows your prompt: low values give looser, more varied images, high values stick closer to the words.

Start on the free Starter plan.

Seed and CFG Scale - FunAI
Seed and CFG Scale - FunAI

Two Numbers That Shape Every Image

The seed decides where the image starts; the guidance scale decides how hard the prompt pulls it along. Lock one and move the other, and AI generation becomes a controlled experiment instead of a lottery.

Open Qwen Image

Seed = starting noise

A whole number that fixes the random noise the model begins from, which makes results repeatable.

CFG = prompt strength

A multiplier on the difference between the prompted and unprompted predictions at every step.

Change one at a time

Fix the seed to test a setting, or fix the settings to explore new seeds. Changing both hides what helped.

Models with seed and guidance inputs on FunAI

Every model page lists its inputs. These are good places to practise, because they expose the seed, the guidance scale or both.

  • •
    Qwen Image:Text-to-image with Seed, Guidance scale, Negative Prompt and inference-step inputs.
  • •
    Stable Diffusion 3.5 Large:A Multimodal Diffusion Transformer with Seed and Guidance scale (CFG) inputs.
  • •
    HiDream I1 Full:Open-source text-to-image model with Seed, Guidance scale and LoRA inputs.
  • •
    FLUX.1 [dev]:A 12 billion parameter flow transformer with Seed and a distilled Guidance scale.
  • •
    Wan 2.2:Text-to-video with Seed and Guidance scale inputs, so you can test prompt changes on the same starting noise.
  • •
    Kling Video v3 Pro:Text-to-video with a CFG Scale that controls how closely the video follows the prompt.
Browse all image models
Seed and CFG Scale Models

Seed and CFG Scale explained

Diffusion models create an image by starting from pure random noise and removing it step by step until a picture appears. That noise is not truly random: it comes from a pseudo-random number generator, and the seed is the number that initialises it. Same seed in, same noise out. Keep everything else fixed and the same seed reproduces the same image; change only the seed and you get a new variation of the same idea. Leave the seed empty and most models pick a random one for each run.

Classifier-free guidance (CFG) is how a model is pushed toward your prompt. At each denoising step the model makes two predictions, one that sees your prompt and one that does not (or that sees your negative prompt), and moves the image along the difference between them, multiplied by the guidance scale. A low scale barely pushes, so images look natural but may ignore parts of the prompt. A very high scale follows the words aggressively and can produce oversaturated colour, harsh contrast and artefacts. Some distilled models, such as FLUX.1 [dev], use a built-in guidance value instead of running the two predictions, so their guidance numbers sit on a different scale.

Fix the seed when you want to compare: try two wordings of a prompt, two guidance values or a LoRA on and off, and any change you see comes from that one difference. Raise the guidance scale when the model ignores an important detail; lower it when results look overcooked. Seeds belong to a specific model and version, so a seed that gave a great image on one model means nothing on another.

Tips for seeds and guidance

  • A seed reproduces a result only with the identical model version, prompt, size, steps and settings; change any of them and the image changes.
  • Move guidance in small steps, around 0.5 to 1 at a time, rather than jumping from 3 to 15.
  • If colours burn out or faces look plastic, lower the guidance before rewriting the prompt.
  • No seed number is better than another in general; a seed is only good for a particular prompt on a particular model.
  • Inference steps are a separate setting: more steps refine detail up to a point, but do not change what the seed and guidance decide.
  • Video models such as Wan 2.2 and Veo 3.1 take a seed too, which helps when testing prompt changes on a single shot.

How to Tune Seed and CFG on FunAI

Feature visual preview 1
Feature visual preview 2
Feature visual preview 3
Feature visual preview 4
01

Open a model that exposes both

Qwen Image, Stable Diffusion 3.5 Large and HiDream I1 Full all show a Seed input and a Guidance scale (CFG) input on their model pages.

02

Explore with a few seeds

Keep the prompt and settings the same and run it with several seed numbers you choose, for example 1, 2 and 3. Pick the composition you like best.

03

Lock the seed, tune the guidance

Keep that seed and change the guidance scale in small steps. Watch whether the prompt is followed more closely or the image starts to look harsh.

04

Refine the prompt on the same seed

With seed and guidance fixed, edit the wording. Because the starting noise is identical, you can see exactly what each word changes.

No public creations to show yet.

Everything you need to know