AI Image

AI Image Settings Explained: Seed, Steps, CFG and Samplers

Learn what seed, inference steps, CFG guidance, samplers and schedulers do in AI image generation, why settings change results, and how to tune them.

Approximately 4 min read

An image-generation workflow becomes easier to tune once each control has a job. The seed selects a starting noise realization. The scheduler and step count determine how the generation process travels through noise levels. Guidance changes how strongly a compatible pipeline follows its conditioning. Resolution changes the size of the problem.

Start with the model’s own example settings, generate a baseline, and change one control at a time. A setting that helps one model can be ineffective on another, especially when moving from older Stable Diffusion pipelines to distilled or flow-matching models.

What the main controls change

Control What it changes Useful first test
Seed Initial random state Explore several seeds to find a composition
Steps Number of numerical updates Compare a short, recommended and longer run
Scheduler or sampler Numerical trajectory through noise levels Compare supported choices at a recorded step count
CFG guidance Conditional guidance in compatible pipelines Test a small range around the model recommendation
Width and height Output geometry and latent tensor shape Start at a supported native resolution
Denoising strength How far an image-to-image process departs from its input Increase gradually when the source image is too restrictive

Seeds are useful for comparisons

A seed initializes a random-number generator; it does not identify a finished image. With the same model, pipeline, prompt, dimensions and software environment, reusing the seed helps compare a change without also changing the starting noise.

Changing resolution changes the noise tensor’s shape. Changing a scheduler changes how that noise is processed. Switching a LoRA, VAE, precision setting or model revision can change the result too. Save these alongside the seed when keeping an image you may want to revisit.

For composition, explore seeds first. Once the framing is promising, keep that seed while testing steps or guidance. Otherwise a better composition can be mistaken for a better parameter value. Diffusers’ reproducibility guide explains how generators and execution environments affect repeatability.

Steps and schedulers belong together

Each step updates the current latent according to the scheduler’s numerical rule. More steps generally mean more work, but image quality can plateau. A distilled model designed for a few steps should be evaluated near its intended operating range rather than given an old 30-step preset.

Names such as DDIM, Euler and DPM++ refer to different sampling methods or implementations. A scheduler also determines the sequence of noise levels, sometimes described using sigmas. Karras schedules arrange those levels in a particular way. The model remains the trained predictor; changing the scheduler changes how that predictor is used.

Compare schedulers with their actual settings recorded. Equal step counts can be a useful control, while an equal-time comparison answers a different question: which setting gives the best result within your time budget?

Guidance and negative prompts

Classifier-free guidance combines conditional and unconditional predictions. Increasing its scale can strengthen prompt adherence, but pushing it too high can produce harsh contrast, oversaturation or distorted detail. Some pipelines use different guidance mechanisms, and distilled models may expect little or no ordinary CFG.

A negative prompt supplies additional conditioning in pipelines that support it. Long lists of unwanted features can change an image in surprising ways. Begin with a short description of the actual problem and check whether the pipeline uses negative conditioning at all.

Image-to-image denoising strength answers another question: how much freedom should the process have to change the input image? Low strength usually preserves more of the starting structure. Higher strength permits larger changes. Treat guidance and denoising strength as separate controls.

A small tuning experiment

Use a prompt with visible details you can check, such as: “A ceramic red teapot on a wooden table, one handle, one spout, soft side lighting.” Choose a seed, a supported resolution and the model’s recommended scheduler.

Round Keep fixed Change Inspect
1 Prompt, model, dimensions, scheduler, guidance Seed Composition and object arrangement
2 Everything including the chosen seed Steps Detail, unfinished areas and runtime
3 Seed, steps and scheduler Guidance Prompt details, colour and distortion
4 Prompt, seed and dimensions Scheduler, with steps recorded Quality and generation time

Keep the images beside their settings. Check the same features in each: number of handles, position of the spout, edge quality and lighting. This gives a parameter change a visible purpose instead of turning tuning into a search for a universally “best” number.

Save a usable record

Record model and revision, prompt, negative prompt, seed, scheduler, steps, guidance, dimensions, LoRAs, VAE and relevant software versions. For image-to-image work, keep the input and denoising strength as well.

When an image is nearly right, make a small controlled change. When the composition is wrong, a new seed or clearer prompt may be more productive. When the image looks unfinished, check the model’s intended step range before increasing everything at once.

Sources and further reading

Continue reading

More in AI Image ↗