Text to image · Prompt guide
AI image generator from text: say more of the picture
A text-to-image model turns a written prompt into a new picture, and the result depends on how much of the picture your prompt decides. This guide explains how the technique works, gives you a six-layer prompt builder, lists the fixes for common flaws, and sets out a test for choosing between the image models in ZeroTwo.
This page shows no generated samples or speed rankings, so the checks are written for you to run.
Anatomy of a prompt
- SubjectStops the model inventing what you did not say.
- StylePhotograph, oil painting, cel-shaded anime, poster.
- CompositionWhere things sit in the frame and from what angle.
- LightingThe gap between a flat image and a cinematic one.
- CameraLens, distance and depth of field, even for illustration.
- MediumThe surface the image looks made on.
How a diffusion model turns words into pixels
Three ideas explain most of what you see when you generate an image, including why prompts behave the way they do.
Training teaches noise removal
A diffusion model is shown real images that have had noise added in small steps until they are static. At each step it learns to predict what was added so it can take it back out.
A text encoder steers each step
Your prompt is converted into numbers by a text encoder. Those numbers push every denoising step toward the words you wrote, which is why specific words change the picture.
Generation starts from noise
To make a new image the model starts from random noise and removes it step by step. Different starting noise gives a different picture, so the same prompt can return different images. Many current models do this in a compressed latent space rather than on full-size pixels.
What that means for your prompt
- Anything you leave out is filled in by the model, so name what matters.
- Specific, concrete words steer more strongly than general praise such as “beautiful”.
- Judge a prompt on several runs, not one, because each run starts from different noise.
Further reading: Denoising Diffusion Probabilistic Models (Ho, Jain and Abbeel, 2020); SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (Podell et al., 2023); Imagen 3 (Imagen-Team-Google, 2024).
Build a prompt one layer at a time
Fill one short line per layer and the assembled prompt updates below. It starts with a worked example, a fox in snow; replace each line with your own subject.
Assembled prompt
Subject: A red fox sitting in fresh snow, mid-winter, mid-distance. Style: Photorealistic wildlife editorial. Composition: Rule of thirds, fox on the right third, snow drift in the foreground. Lighting: Late-afternoon golden hour, soft side-light, low contrast. Camera: 85mm portrait lens, f/2.8, shallow depth of field. Medium: 35mm film grain, subtle warmth, scanned print.
Copy the text into the ZeroTwo chat. Nothing typed here leaves your browser.
Open ZeroTwo and paste itSix prompts to copy, and what to check in each result
Each template uses all six layers. The note under it says what to look for, so you judge the image against the prompt and not against a feeling.
Portrait
Subject: a woman in her sixties, silver hair pulled back, soft smile lines. Style: photorealistic editorial portrait. Composition: tight bust shot, slight three-quarter turn, eyes meeting camera. Lighting: north-window soft light, single key, no rim. Camera: 85mm, f/1.8, shallow depth of field. Medium: medium-format film, neutral grade.
Check: Eyes and skin texture, and whether the light looks like one soft window.
Landscape
Subject: a windswept black-sand beach with basalt sea stacks. Style: cinematic photoreal. Composition: wide angle, low horizon, long-exposure surf. Lighting: blue-hour twilight, cool ambient with a single warm lamp on a distant figure. Camera: 16mm wide, f/11, 30-second exposure. Medium: digital, slight clarity boost.
Check: Whether the one warm light and the distant figure both appear.
Product
Subject: a brushed-titanium chronograph wristwatch on a folded white linen napkin. Style: luxury catalog photography. Composition: top-down 45-degree angle, watch in the lower-right third, generous negative space. Lighting: soft box from upper left, gentle bounce card on the right, no specular blowout. Camera: 100mm macro, f/8. Medium: studio digital, retouched.
Check: Dial detail, the folded napkin, and any blown-out highlights on the metal.
Illustration
Subject: a cat astronaut floating outside a tea-shop spaceship. Style: cel-shaded anime illustration with a warm palette. Composition: centered subject, helmet glass reflecting Earth, asymmetric stars. Lighting: hard rim from the left, warm under-fill. Camera: not applicable. Medium: digital paint with visible brushwork in the highlights.
Check: Whether the helmet reflection and the tea shop are both present.
Abstract
Subject: an architectural abstraction of a midcentury staircase. Style: high-contrast geometric photography. Composition: looking straight up the stairwell, perfect symmetry, vanishing point at center. Lighting: top-down hard daylight. Camera: 24mm, f/16. Medium: black-and-white film, deep blacks, paper-white highlights.
Check: Symmetry, the vanishing point, and true black and white with no color cast.
Typography
Subject: a vintage coffee-shop menu chalkboard listing "House Espresso", "Cortado", "Latte" and "Filter". Style: hand-lettered with art-deco flourishes. Composition: centered title, three columns, ornamental borders. Lighting: warm lamp light from above. Camera: not applicable. Medium: chalk on slate, photographed straight on.
Check: Spelling of all four menu words, letter by letter.
Why an image looks off, and the next change to try
Change one thing per run so you can tell which edit helped.
| What you see | Likely cause | Try this |
|---|---|---|
| Skin looks plastic | The prompt leaves finish unspecified, so the model defaults to a smooth, retouched look. | Name a film stock or print process, a shallow aperture, and ask for visible skin texture and no beauty retouching. |
| Hands or faces are wrong | Small structures are where errors are easiest to see, and a crowded prompt leaves less attention for them. | Describe the hands plainly, simplify the pose, then generate several variations and keep the cleanest. |
| Text in the image is misspelled | Letters are drawn as shapes, so exact spelling is not guaranteed. | Put the exact string in quotes, keep it short, and test the same prompt on more than one model. |
| Lighting looks flat | With no lighting named, the model picks a neutral, even default. | Name time of day, direction, hardness and color temperature in one phrase. |
| A requested detail is missing | Long prompts bury details, and the model weighs some words more than others. | Move the missing detail to the front of its layer, cut decoration, and check each detail against a written list. |
| The style drifts between runs | Each run starts from different random noise. | Repeat the same style words exactly, and use one variation as the base for the next prompt. |
The image models in ZeroTwo, and how to pick one yourself
ZeroTwo lists image models from OpenAI, Google, Black Forest Labs, Alibaba and xAI. We have not measured them against each other in a way you could reproduce, so there is no ranking, no latency column and no resolution table here. If you are choosing between them to buy or commit to, the guide to choosing an image model for your prompt covers that decision; if you are starting from an existing picture, see AI image editing. To carry the visuals into a campaign, see the creative AI workflow, or turn the direction into a shot list with the AI-powered video creation platforms guide.
GPT Image 1.5Also edits images
OpenAI
OpenAI's newest GPT Image entry in the list.
GPT Image 1Also edits images
OpenAI
OpenAI image generation, also used for editing.
GPT Image 1 Mini
OpenAI
A faster, lighter OpenAI image model.
Imagen 4
Google
Google's Imagen image generation model.
Nano BananaAlso edits images
Google Gemini
Conversational image generation and editing.
Nano Banana ProAlso edits images
Google Gemini 3 Pro
Image generation and editing on Gemini 3 Pro.
Flux Pro
Black Forest Labs, via FAL
The FLUX Pro model served through FAL.
Flux Pro 2Also edits images
Black Forest Labs, via FAL
The newer FLUX Pro model, also used for editing.
Qwen
Alibaba
Image generation from Alibaba's Qwen family.
Grok 2
xAI
Image generation from xAI's Grok.
A five-check test you can run in ten minutes
Use the same prompts on every model you want to compare, write down the model name and date, and record counts rather than impressions.
Exact text
Prompt a sign with four specific words. Run it four times per model and count how many runs spell every word right.
Anatomy
Prompt a person holding an object in both hands. Count the runs with correct fingers and a sensible grip.
Prompt adherence
Write five required details before you generate. Tick each one per image and score the model out of five.
Style match
Prompt one named style, such as watercolor or film photograph, and judge each result against a reference you chose beforehand.
Usable per batch
Generate four variations of one prompt and count how many you would use without editing.
What image generation costs in ZeroTwo
Image generation is metered. Chat on fast models is unlimited on Plus, but every image, video and deep-research run draws credits from the same monthly pool, and heavier models and longer tasks use more of them. We do not publish a credits-per-image figure, so treat the credit numbers as a budget to test against, not a count of images.
Free
$0/mo
- Limited image generations
- +20 bonus credits every day you log in
- No card required
Plus
$14.99/mo
- Image and video generation
- 1,650 monthly credits
- 1,000 free Auto requests / month
Pro
$29.99/mo
- 3,500 monthly credits
- 2,500 free Auto requests / month
- Image runs draw from the same credit pool
Read from the same plan catalog as the pricing page, which also lists Pro 2x, Ultra and Business.
Questions about text-to-image generation
What is an AI image generator from text?
It is a model, usually a diffusion model, that takes a written prompt and produces a new image that matches it. It has learned from large sets of images paired with captions how words relate to visual features such as lighting, style and composition.
Which text-to-image model should I use?
There is no single answer, and this page does not rank models. Run the same prompt on two or three models from ZeroTwo's list and score the results with the five checks above. The right model depends on whether you need exact text, clean anatomy, a particular style or speed.
Can I use generated images commercially?
It depends on the model provider's terms and on your plan, and those terms change. Read the terms of the model you used before you rely on an image in paid client work, a product or an advertisement. This page is not legal advice.
How long does it take to generate an image?
It varies by model, image size, prompt and how busy the service is. We do not publish timings because we have not measured them in a way you can reproduce. Time a few runs in your own workflow and judge from those.
Why does my image have the wrong hands or garbled text?
Small structures and exact lettering are where diffusion models fail most visibly. Simplify the pose, quote the exact text, generate several variations, and try a second model. The fix table on this page lists the next change to try for each flaw.
Is image generation free in ZeroTwo?
The Free plan includes limited image generations, 100 free credits to start, 20 bonus credits every day you log in and no card requirement. On paid plans every model, image, video and deep-research run draws from one monthly credit pool: $14.99 a month for Plus with 1,650 monthly credits, $29.99 a month for Pro with 3,500 monthly credits. Heavier models and longer tasks use more credits. We do not publish a credits-per-image figure here.
Where can I read how ZeroTwo handles my prompts and images?
In the privacy policy at zerotwo.ai/privacy. This page makes no separate claim about retention or training.
Keep going
Choose an image model for your prompt and your budget.
Start from an image you already have instead of a blank prompt.
A formula and worked examples for describing a picture.
Ready-to-adapt prompts grouped by category.
The feature page for generating and refining images in a chat.
What you can do on a free plan.