КАДР

Seed, CFG, steps, resolution: what AI generation settings mean

Alen founder of KADR Updated 2026-10-01 5 min read

Generation has a few settings that scare beginners with obscure names: seed, CFG, steps, resolution. In fact their logic is simple. Seed controls the repeatability of the result, CFG how strictly the model follows the prompt, steps how much detail is worked out, resolution the size and the price. Understanding these four parameters saves money and nerves, because some of them really affect the result and some work as a placebo. Below we go through each one separately, with a definition in the first line and its practical meaning.

Why bother understanding settings at all

Settings scare people with their names, but their logic is simple, and understanding it saves money. Some parameters really change the result, others barely matter, and some are a pure placebo. Once you know which is which, you stop overpaying for useless «improvements» and start adjusting what actually works.

Let us go through the four main ones: seed, CFG, steps, resolution. If you want more terms after this, there is a full AI generation glossary.

Seed: a number for repeating

A seed is the number from which the starting noise of a generation is born. From that noise the model then develops the picture (how this works is covered in the article on diffusion).

The practical meaning is simple: the same seed plus the same prompt give an identical result. Change the seed and you get a different variant of the same description.

Why you need it:

  • Lock in a success. You got a great shot, note its seed. Now you can tweak the prompt slightly while keeping the base.
  • Browse variants. One prompt with different seeds is a way to get many variations of one idea.
  • Make a series. Similar seeds help keep a consistent style across a set of pictures.

The seed is your control tool. Without it every generation is random; with it you control repeatability.

CFG: how strictly to obey the prompt

CFG (guidance scale) is the parameter that sets how strictly the model follows your text.

  • Low CFG: the model is freer, with more creativity and surprises, but it may drift from the prompt.
  • High CFG: the model sticks to the text rigidly, but at too high a value the picture gets «overcooked»: the colours turn acidic and the image unnatural.
  • The optimum is in the middle. Most good results lie in the mid range.

In practice: touch CFG only if you have a problem. The model keeps ignoring important parts of the prompt? Raise CFG. The picture looks overcooked and oversaturated? Lower it. If everything is fine, leave the default.

Steps: how many passes to develop

Steps are the number of passes in which the model removes noise and finishes the picture.

The logic: more steps mean slightly more worked-out detail, but also slower and more expensive. However, there is a trap: beyond a certain number of steps the difference is barely visible. The model has already developed everything it could, and extra steps just burn time and money without improving the picture.

In practice: in most services steps can be left at the default; a reasonable value is already set. Pushing them to the maximum for the sake of «quality» makes no sense; it is one of the most common misconceptions.

Resolution: the main price lever

Resolution is the size of the picture or video in pixels, and it is the parameter with the biggest effect on price.

For video the difference is especially large. The same clip can cost five times more at high resolution than at low. An example with Seedance 2.0: a 6-second clip costs 69 tokens at 480p and 362 tokens at 1080p.

Hence the main saving rule in all of generation: drafts at low resolution, the final at high. While you are tuning the prompt and testing the idea, you do not need detail; what matters is composition and logic. Those are perfectly visible at low resolution for pennies. At high resolution you make one final, already debugged version.

How resolution affects the cost of a minute of video is calculated in the article on how much a minute of AI video costs.

What really matters and what is a placebo

Let us put it into a clear picture.

Really matters: - Resolution for price and detail (most of all), - seed for repeatability, - CFG for how precisely the prompt is followed.

Barely matters: - steps above a reasonable threshold (extra steps are almost useless).

Placebo: - add-ons in the prompt like «8K», «ultra-detailed», «maximum quality». They hardly change the result and just take up space. More on this in the article on prompts for video.

What a beginner should adjust

A simple plan:

  1. Resolution: low for tests, high for the final. This is the main saving.
  2. Seed: lock it in when you catch a good shot.
  3. CFG: touch it only when there is a problem (it ignores the prompt or looks overcooked).
  4. Steps: leave the default.

That is enough to control the result deliberately and not overpay.

In short

The four main settings in plain words: seed is repeatability, CFG is how strictly the prompt is followed, steps are the development passes, resolution is size and price. Resolution has the biggest effect on money, so do drafts at low resolution and the final at high. The seed locks in a success, touch CFG when there are problems, leave steps as they are. And «8K» add-ons in the prompt are a placebo.

Frequently asked questions

What is a seed in generation?
A seed is the number from which the starting noise for a generation is created. The same seed with the same prompt gives an identical result. Change the seed and you get a different variant of the same description. It is a way either to lock in a good shot or to browse variants.
What is CFG or guidance scale?
CFG shows how strictly the model sticks to your prompt. A low value gives more freedom and creativity but less precision. A high one forces it to follow the text, but too high and the picture becomes overcooked and unnatural. The optimum is usually in the middle.
What are steps in generation?
Steps are the number of passes in which the model removes noise and reveals the picture. More steps mean slightly more detail, but slower and more expensive. Beyond a certain number the difference is barely visible, so there is no point pushing steps to the maximum.
Does resolution affect the generation price?
Yes, a lot. The higher the resolution, the more expensive the generation. For video the difference between low and high resolution can be fivefold. So drafts are best done at low resolution, and high resolution is turned on only for the final.
Which settings really matter and which are a placebo?
Really matter: resolution (for price and detail), seed (for repeatability), CFG (for how precisely the prompt is followed). Barely matters: steps above a reasonable threshold. Pure placebo: add-ons like «8K, ultra-detailed» in the prompt; they hardly change the result.
Which settings should a beginner change and which should they leave alone?
Start with resolution (low for tests, high for the final) and seed (lock in a good shot). Touch CFG only if the result systematically ignores the prompt or looks overcooked. In most services steps can be left at the default.

Settings are visible before generation

On KADR you see the resolution, duration and price in tokens before you start, so it is easy to pick parameters without surprises. Drafts at low resolution cost pennies, the final at high. Pay in rubles, tokens never expire.

Open KADR
Alen
Alen
Founder of KADR. I connected the generation models directly through their APIs, so I explain the settings in terms of what really affects the result and the price.