What is a seed in AI image generation?

A seed is the number fixing the random noise a generation starts from. The same seed, prompt and model reproduce the same image, which is why fixing it is the only way to test what a single word change actually did.

What does denoise strength control?

How much of your source image survives an image-to-image generation. At 0.2 the photograph is essentially intact; at 0.5 composition holds while content is renegotiated; at 0.85 the source is little more than a mood board.

What is the difference between inpainting and outpainting?

Inpainting regenerates a masked region inside the existing frame; outpainting generates new content beyond the original edges. The same operation pointed at different areas — one repairs, the other extends.

What is guidance scale?

How strongly the model is pushed toward your prompt at each denoising step. Too low and it drifts away from your words; too high and the output becomes rigid and over-saturated as it over-corrects.

A small trained add-on that teaches a base model one specific style, character or object without retraining the whole model. It is how open-model users get consistent characters across many images.

Why does aspect ratio need choosing before generating?

The model composes for the frame it is given, arranging a scene horizontally for 16:9 and vertically for 4:5. Cropping afterwards throws away composition it made deliberately.

What is an alpha channel?

The transparency layer in an image file. PNG and WebP support one; JPEG does not, which is why a cut-out subject saved as JPEG gains a solid white background.

Does a seed work in a different model?

No. A seed indexes into one specific model's noise space, so the same number in another model — or often another version of the same model — produces something unrelated. Record the model alongside it.

AI image glossary

The vocabulary an AI image generator expects you to already know

Every term below appears in a real interface or a real prompt. Each is defined in one sentence first, because a definition you have to read twice has failed.

Most difficulty with AI image generation is vocabulary rather than technique. Seed, denoise strength, guidance scale and inpainting each name one specific behaviour, and once the word means something the control becomes obvious. These are the terms worth knowing, in the order you meet them.

How generation works

Diffusion: The technique behind most current image models: trained by adding noise to real images and learning to reverse it, then run backwards from pure noise to generate.
Text to image: Generating a new picture from a written description alone, with no source image. Also called text to photo when the intended output is photographic.
Image to image: Generating a new picture from an existing one plus a prompt, with denoise strength controlling how much of the source survives.
Checkpoint: A saved set of model weights. Different checkpoints of the same architecture produce visibly different output from identical prompts.

Controls and parameters

Seed: The number that determines the starting noise a generation begins from. The same seed with the same prompt and model reproduces the same image.
Denoise strength: In image-to-image, how much of your source image is discarded before regeneration. Low values preserve the original; high values treat it as a loose reference.
Guidance scale: How strongly the model is pushed toward your prompt at each step. Too low and it wanders; too high and the output becomes rigid and over-saturated.
LoRA: A small trained add-on that teaches a base model a specific style, character or object without retraining the whole model.
Negative prompt: A separate field describing what to avoid. Not all endpoints support one — where they do not, everything you write is a positive instruction.
Sampler: The algorithm that steps the denoising process. Different samplers converge at different speeds and produce slightly different results from the same seed.

Editing operations

Inpainting: Regenerating a masked region inside an image while leaving everything outside the mask untouched. The basis of object removal and targeted editing.
Outpainting: Generating new image content beyond the original frame edges, used to extend a photograph to a wider aspect ratio without cropping.

Files and output

Aspect ratio: The proportional relationship between an image’s width and height, such as 1:1, 4:5 or 16:9. Chosen before generating, because the model composes for the frame.
Upscaling: Enlarging an image beyond its original pixel dimensions. AI upscaling reconstructs plausible detail rather than interpolating between existing pixels.
Alpha channel: The transparency layer in an image. PNG and WebP support one; JPEG does not, which is why a transparent cutout saved as JPEG gains a solid background.
EXIF: Metadata embedded in a photograph recording camera, lens, exposure, date and often GPS coordinates. Stripped whenever a file is re-encoded.

How this glossary works

Terms: Only ones you meet in a real interface
Definitions: One sentence first, always
Own page: Only where a paragraph is not enough
Schema: DefinedTermSet, with members linked
Updated: When model behaviour changes
Cost: Free, no account

In practice

Most AI image difficulty is vocabulary

Seed, denoise strength, guidance scale and inpainting each name exactly one behaviour. Once the word means something, the control it describes becomes obvious — and until then the interface looks like a wall of arbitrary sliders.

That is why definitions come first here and explanation second. A definition you have to read twice has failed at the only job it had.

An open technical reference book on warm paper with a magnifier resting on a page of diagrams

Three ways in

Where to start

How images are made

Diffusion, text to image, image to image and checkpoints. The mechanics that explain why the same prompt gives different results.

Start with diffusion, then seed.
Text to image and image to image are different modes.
Checkpoints change output more than prompts do.

Detail

Glossary questions

The vocabulary questions that come up most.

Why the same prompt gives different images

The seed was not fixed. Generation starts from random noise, and the starting point affects the result more than most prompt edits.

What denoise strength does

Sets how much of a source image survives an image-to-image generation. It is the single most important control in that mode.

The difference between inpainting and outpainting

Inpainting regenerates inside the frame; outpainting generates beyond its edges. Same operation, different area.

Whether a seed works across models

No. A seed indexes into one model's noise space, so the same number elsewhere produces something unrelated.

Why aspect ratio should be chosen first

The model composes for the frame it is given, so cropping afterwards discards arrangement it made deliberately.

An AI-generated product photograph on a white sweep

Apply the vocabulary

the main generatorText to finished image.

Image to ImageDenoise strength in practice.

AI Photo EditorInpainting in practice.

AI Image UpscalerUpscaling in practice.

AI prompt generatorBuild a structured prompt.

all photo and art stylesStyle as a control.

Keep exploring

Elsewhere on LaFoto

Every term here describes something one of these does.

generate from a descriptionThe main generator — text to finished image.Open Prompt Gallery120 images with the exact prompts that made them.Open start from a photoStart from a picture you already have.Open edit part of a photoChange part of a photo, keep the rest.Open enlarge an imageEnlarge to 4K without the smearing.Open enhance a photoExposure, noise and old-photo restoration.Open masked object removalMasked removal that rebuilds the background.Open Blur BackgroundDepth-aware, not a uniform filter.Open Colorize PhotoPlausible colour for black and white.Open AI Logo GeneratorMarks that still read at 16 pixels.Open design a tattooDesigns that survive being tattooed.Open generate illustrationIllustration and painting, not photographs.Open generate a headshotStudio-grade portraits without a studio.Open passport photo requirementsThe exact specification, and how to meet it.Open interior visualisationRestyle a room you can photograph.Open terms explainedSeed, denoise, inpainting, explained.Open

Vocabulary

Questions about AI image terms

What is a seed in AI image generation?: A seed is the number fixing the random noise a generation starts from. The same seed, prompt and model reproduce the same image, which is why fixing it is the only way to test what a single word change actually did.
What does denoise strength control?: How much of your source image survives an image-to-image generation. At 0.2 the photograph is essentially intact; at 0.5 composition holds while content is renegotiated; at 0.85 the source is little more than a mood board.
What is the difference between inpainting and outpainting?: Inpainting regenerates a masked region inside the existing frame; outpainting generates new content beyond the original edges. The same operation pointed at different areas — one repairs, the other extends.
What is guidance scale?: How strongly the model is pushed toward your prompt at each denoising step. Too low and it drifts away from your words; too high and the output becomes rigid and over-saturated as it over-corrects.
What is a LoRA?: A small trained add-on that teaches a base model one specific style, character or object without retraining the whole model. It is how open-model users get consistent characters across many images.
Why does aspect ratio need choosing before generating?: The model composes for the frame it is given, arranging a scene horizontally for 16:9 and vertically for 4:5. Cropping afterwards throws away composition it made deliberately.
What is an alpha channel?: The transparency layer in an image file. PNG and WebP support one; JPEG does not, which is why a cut-out subject saved as JPEG gains a solid white background.
Does a seed work in a different model?: No. A seed indexes into one specific model's noise space, so the same number in another model — or often another version of the same model — produces something unrelated. Record the model alongside it.

Start creating today

Generate your first image with the best AI image generator.

Turn a sentence into a finished, photorealistic image in seconds — then refine every detail. No setup, no Discord, no GPU.

Try these controls yourself Get Started

Join 4,200+ creators using LaFoto