What is the difference between an open and a closed model?

Open weights can be downloaded and run on your own hardware, so you can pin a version permanently, generate without a per-image cost and keep source material off third-party servers. Closed models are reached through an API or an app, require no infrastructure, and change on the provider’s schedule rather than yours.

Why do these pages not list prices or version numbers?

Because both change faster than a directory page gets revised, and a stale number reads as a lie rather than as staleness. What is written here is the part that persists — who makes it, how you reach it, whether the weights are open, and what it is built to be good at.

Is LaFoto affiliated with any of these models?

No. We do not resell any of them and have no commercial relationship with their makers. The one place we appear in the first person is the FLUX entry, where we say which model generated this site’s own imagery — a claim the gallery substantiates card by card.

Is a newer model always better than an older one?

On raw output quality, usually. On everything surrounding it — community tooling, style adapters, licence clarity — often not. Stable Diffusion is the clearest example of an older model kept relevant by its ecosystem rather than by its results.

Which model should I use to edit a photo I already have?

One built for editing rather than generation. The distinguishing capability is holding a subject recognisably constant across successive changes, which most text-to-image models lose after one or two edits.

Do any of them render text reliably?

Several handle short strings well enough to design around, and none is a substitute for setting type. Generate the picture and add the words in a design tool — generated lettering cannot be edited, spell-checked or translated afterwards.

AI Image Models Explained — A Plain Directory

These models are not competing at the same thing. Midjourney imposes an aesthetic, FLUX takes instruction literally, Nano Banana changes a photograph without losing the person in it, Stable Diffusion hands you the parts, Recraft outputs editable vectors, Seedream sets Chinese and English type, and gpt-image-1 parses a complicated sentence. Pick by which of those sentences describes your problem.

The directory

GoogleNano BananaGoogle’s conversational image editor. Keeps a face consistent across edits.
Stability AIStable DiffusionThe open model the whole ecosystem grew from. LoRAs, ControlNet, checkpoints.
Black Forest LabsFLUX.1Built by ex-Stable Diffusion researchers. Reads prompts literally, renders text.
Midjourney, Inc.MidjourneyThe one with a house style. Opinionated, closed, subscription-only.
RecraftRecraftThe one that outputs real vectors. Built for brand and design work.
ByteDanceSeedreamByteDance’s model. Strong bilingual text rendering and high native resolution.
OpenAIgpt-image-1The model behind ChatGPT’s images. Follows instructions, renders text well.

How to choose between them

The question that decides most of it is whether you are making an image or changing one. Making an image from nothing rewards a model with composition and prompt adherence; changing a photograph rewards a model that leaves alone everything you did not mention. Those are close to opposite requirements, and a tool built for one is usually mediocre at the other.

The second question is whether you need the result to be reproducible. Closed hosted models change underneath you, and a look you rely on commercially can stop being available on a schedule you do not control. Open weights fix that at the cost of running the infrastructure yourself.

The third is output format, and it is the one people discover too late. If the file has to open in a vector editor or print at two metres, that constraint eliminates most of this list before any question of quality arises.

What this directory is

Entries: Seven models
Ranked: No — described
Affiliation: None with any of them
Pricing: Deliberately not listed
Versions: Not pinned — they move too fast
Updated: August 2026

How to read a model comparison

Almost every ranking is answering a question you did not ask

Lists of the best image model are usually built by generating the same handful of prompts through each one and judging the results. That measures how each model handles those prompts, which is only useful if your work resembles them, and a set of test prompts chosen to look good in a blog post rarely does.

The more useful question is structural: does this model make images or change them, does it have an aesthetic opinion, can you run it yourself, and what comes out of the other end — pixels or curves. Those properties do not shift with the next release, and they eliminate most of the list before quality enters the discussion.

Seven small framed photographic prints arranged in a grid on a pale studio wall

Three questions that narrow the field fast

Choose by constraint before you choose by quality

The first fork, and the widest one

Generating an image from a description and editing a photograph you already have are close to opposite problems. One rewards composition and prompt adherence; the other rewards leaving untouched everything you did not mention.

Models built for editing hold a subject's identity across successive changes. Models built for generation give you control over framing, optics and style from nothing. Very few are excellent at both.

Editing an existing photo: Nano Banana.
Making one from nothing: FLUX, Midjourney.
Both, conversationally: gpt-image-1.

An AI-edited product scene with a replaced background

Model questions

The things people ask before choosing one

Short answers, with the longer version on each model’s own page.

Why do the same prompt and model give different images?

Because the starting noise differs. A seed fixes that noise, and prompt plus seed plus model reproduces an image exactly. Change any one of the three and you get something related but not identical.

Is a newer model always better?

On raw output, usually. On everything around it — community tooling, style adapters, licence clarity, price stability — often not. Stable Diffusion is the clearest case of an older model kept alive by its ecosystem.

Can I move my prompts between models?

The subject and scene transfer; the parameters and syntax rarely do. Models differ in how they read weighting, negatives and style tokens, and a prompt tuned hard for one is usually over-specified for another.

What does open weights actually get me?

The ability to run it locally, pin a version forever, generate without a per-image cost, and keep source material off third-party infrastructure. It costs you the hardware and the maintenance.

Which one renders text properly?

Several handle short strings now, and none is a substitute for setting type. Generate the picture, add the words in a design tool, and you avoid the whole category of problem.

Does the model matter more than the prompt?

Below a certain standard of prompt, no. A specific prompt through a mid-tier model beats a vague one through the best model available, which is why the prompt guidance here is longer than this directory.

An AI-generated food photograph in daylight

Where to go from the directory

Prompt Gallery120 images with the model, prompt and seed on each.

the glossarySeed, sampler, LoRA, guidance scale, defined.

how LaFoto comparesSide-by-side against eight named alternatives.

the ranked comparisonThe ranked commercial view, kept separate.

structure a promptBuild a structured prompt for any of them.

style as a controlStyle as a control rather than a model choice.

Keep exploring

Elsewhere on LaFoto

The directory describes the tools; these are the things you do with them.

generate from a descriptionThe main generator — text to finished image.Open 120 images and their prompts120 images with the exact prompts that made them.Open Image to ImageStart from a picture you already have.Open AI Photo EditorChange part of a photo, keep the rest.Open upscale to 4KEnlarge to 4K without the smearing.Open enhance a photoExposure, noise and old-photo restoration.Open remove something from a photoMasked removal that rebuilds the background.Open depth-aware blurDepth-aware, not a uniform filter.Open add colour to an old photoPlausible colour for black and white.Open AI Logo GeneratorMarks that still read at 16 pixels.Open tattoo flash designsDesigns that survive being tattooed.Open generate illustrationIllustration and painting, not photographs.Open studio headshotsStudio-grade portraits without a studio.Open the photo specificationThe exact specification, and how to meet it.Open interior visualisationRestyle a room you can photograph.Open terms explainedSeed, denoise, inpainting, explained.Open

Questions about AI image models

Which AI image model is best?: There is no single answer, which is why this directory describes rather than ranks. Midjourney imposes an aesthetic, FLUX takes instruction literally, Nano Banana edits a photograph without losing the person in it, Stable Diffusion hands you the parts, Recraft outputs vectors, Seedream sets Chinese and English type, and gpt-image-1 parses a complicated sentence. Pick by which of those describes your problem.
What is the difference between an open and a closed model?: Open weights can be downloaded and run on your own hardware, so you can pin a version permanently, generate without a per-image cost and keep source material off third-party servers. Closed models are reached through an API or an app, require no infrastructure, and change on the provider’s schedule rather than yours.
Why do these pages not list prices or version numbers?: Because both change faster than a directory page gets revised, and a stale number reads as a lie rather than as staleness. What is written here is the part that persists — who makes it, how you reach it, whether the weights are open, and what it is built to be good at.
Is LaFoto affiliated with any of these models?: No. We do not resell any of them and have no commercial relationship with their makers. The one place we appear in the first person is the FLUX entry, where we say which model generated this site’s own imagery — a claim the gallery substantiates card by card.
Can I move a prompt from one model to another?: The subject and the scene transfer; the parameters and syntax rarely do. Models differ in how they read weighting, negatives and style tokens, and a prompt tuned hard for one is usually over-specified for another.
Is a newer model always better than an older one?: On raw output quality, usually. On everything surrounding it — community tooling, style adapters, licence clarity — often not. Stable Diffusion is the clearest example of an older model kept relevant by its ecosystem rather than by its results.
Which model should I use to edit a photo I already have?: One built for editing rather than generation. The distinguishing capability is holding a subject recognisably constant across successive changes, which most text-to-image models lose after one or two edits.
Do any of them render text reliably?: Several handle short strings well enough to design around, and none is a substitute for setting type. Generate the picture and add the words in a design tool — generated lettering cannot be edited, spell-checked or translated afterwards.

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.

Generate an image with LaFoto Get Started

Join 4,200+ creators using LaFoto

AI image models, described rather than ranked