What people actually ask about generating AI images
Long answers, not tips lists. Every one of these is written against the three models we publish with — OpenAI gpt-image-2, FLUX 2 Pro and Nano Banana 2 — so it says what is actually available rather than describing a slider panel that no longer exists.
Generation settings
What seed, aspect ratio, guidance and negative prompts actually control — and which of them exist on the models we publish with.
What do seed, aspect ratio, guidance and steps actually do?
Seed fixes the random starting point so a prompt is reproducible. Aspect ratio changes framing, not just crop. Guidance (CFG) sets how literally the model follows your words. Steps control how long it refines. Modern hosted models expose far fewer of these than older Stable Diffusion UIs did — often only size and a reference image.
read →Which aspect ratio should I use, and how does it change the image?
Aspect ratio determines what the model composes, not how it's cropped afterwards. Portrait frames push subject-dominant compositions, wide frames push environmental ones, square sits between. Choose it before you write the prompt, and match your camera language to it.
read →Do negative prompts work, and what should go in one?
Negative prompts work on models that support a separate negative conditioning channel — mainly Stable Diffusion and FLUX-family models. They exclude visual concepts, not facts. Long copy-pasted 'bad quality, worst quality' lists mostly waste conditioning; short, specific exclusions of things the model is actually producing work far better.
read →Writing the prompt
The order words go in, which camera and lighting terms actually move the image, and why a prompt stops obeying you past a certain length.
How should a prompt be structured, and does word order matter?
A prompt that behaves predictably names one subject first, then what it is doing, then where, then the light, then the lens or medium, then the style. Order matters because everything you write competes for the same limited attention — what comes first survives, what comes last gets diluted.
read →Which camera and lighting terms actually change an AI image?
Focal length, aperture, light direction and time of day change the image reliably because they correlate with real photographic patterns in the training data. Brand names, sensor models and invented gear specs mostly do nothing.
read →Why does the model ignore parts of my prompt?
Ignored instructions are almost always conflicts, counts, or spatial relationships — three things current models handle badly — rather than the model failing to read you. The fix is to remove the competing clause, not to repeat the one being dropped.
read →When the image goes wrong
Hands, faces, text, extra limbs. Why these specific failures happen, and the prompt-side fixes that hold up.
Why does AI get hands and faces wrong, and how do I fix it?
Hands fail because they're small in frame, extremely variable in pose, and self-occluding — the model has less usable signal per pixel than for any other body part. The reliable fixes are compositional: change what the hands are doing, change the framing, or move them out of the frame. Newer models are markedly better but not immune.
read →Why can't AI spell, and how do I get readable text in an image?
Older diffusion models had no character-level representation of text, so they produced letterform-shaped texture instead of words. Current models are far better at short strings but still degrade past a few words. The workable approach is short text, explicitly quoted, on a plausible surface — or compositing real type afterwards.
read →Keeping things consistent
Same character, same style, across a whole set — the part that trips people up once they move past single images.
How do I keep the same character across multiple images?
Text prompts alone won't hold a face across images — a description defines a type, not a person. What works is anchoring on a reference image and editing from it, keeping the character description byte-identical between prompts, and accepting a consistency budget: face and hair hold well, exact clothing details drift.
read →How do I keep one consistent style across a whole set of images?
Style consistency is more achievable than character consistency because style lives in describable attributes. Build a fixed style block — medium, lighting, palette, camera, grade — reuse it verbatim across every prompt, keep one model for the whole set, and vary only the subject clause.
read →Choosing a model
gpt-image-2, FLUX 2 Pro and Nano Banana 2 behave differently on the same words. Where each one is the right call, and when a reference image beats a rewrite.
gpt-image-2, FLUX 2 Pro or Nano Banana 2 — which should I use?
Pick by failure mode, not by leaderboard. gpt-image-2 follows complicated instructions and renders text best, FLUX 2 Pro gives the most photographic surfaces and light, Nano Banana 2 is the fastest to iterate with and the strongest at editing an existing image.
read →When should I use a reference image instead of a better prompt?
Use a reference when the thing you want is easier to show than to describe — a specific face, an exact layout, a palette, an existing frame you only want to change slightly. Use words when you want a new idea rather than a variation of an old one.
read →Working backwards from an image
You have the picture and not the prompt. What can actually be recovered from a finished image, what has to be guessed, and how tag-style descriptions differ from sentences.
How do you get the prompt from an AI image?
Sometimes the prompt is literally inside the file: PNGs written by Automatic1111, ComfyUI and similar local tools store the prompt, negative prompt, seed, sampler and model in the image metadata, and you can read it in seconds. When it is not there — which covers almost everything downloaded from a website, a phone, or any hosted model like gpt-image-2 or Nano Banana 2 — nothing recovers the original wording, and the honest job becomes describing the image well enough that a model reproduces it.
read →What does "image to prompt" actually mean, and do those tools work?
"Image to prompt" describes two different things sold under one name: reading a prompt that a local tool embedded in a PNG, and having a vision model write a fresh description of a picture. The first is exact and works only on files that still carry their metadata. The second always produces something, never produces the original wording, and its quality depends entirely on how well it describes lighting, camera and style rather than just naming objects.
read →Should an image-to-prompt result be tags or a sentence?
Tag lists ("1girl, red dress, rim light, 85mm") come from the booru-trained anime lineage and still work well on those checkpoints. Today's hosted models — gpt-image-2, FLUX 2 Pro, Nano Banana 2 — were trained on natural captions and read sentences far more accurately, especially for spatial relationships and anything involving more than one subject. Tags remain useful as a checklist and as the vocabulary layer, not as the final prompt.
read →After the image exists
Resolution and print, how many attempts is normal, and what you are actually allowed to do with the file.
What resolution do AI images come out at, and does upscaling help?
Current models generate around one to two megapixels natively, which is fine on screen and short for print. Upscaling adds pixels and can invent plausible detail, but it also enlarges every existing flaw — fix the image first, then enlarge.
read →Who owns an AI-generated image, and can I use it commercially?
Usage rights come from the provider's terms, and they generally let you use what you generate, including commercially. Copyright ownership is a separate question: in the United States, purely machine-generated output is not registrable, and other jurisdictions differ.
read →How many tries should one good AI image take?
For a simple subject, expect a usable frame within a handful of generations. For a specific brief, ten to thirty is ordinary. What separates a long session from a wasted one is changing exactly one thing at a time and keeping the rejects.
read →Have the image but not the prompt?
Working backwards is its own skill: read the file's metadata first, and when that comes up empty, reconstruct the prompt component by component. We are also building an image-to-prompt engine on top of this library's labelled prompts.
Looking for measured data instead?
These pages are how-to. For what our own published prompts actually measure out to — length, structure, vocabulary, cross-model behaviour — see reports.
Looking for a specific prompt instead?
Every prompt page carries its own FAQ about that exact image — what the wording is doing and how to change it. These pages are the general layer underneath.