Free tool · Runs in your browser · Rules verified 2026-10-17
Type a rough idea, pick the shape, light, and mood, and get three prompts ready to paste: one written the way ChatGPT Images follows a brief, one as the scene description Google recommends for Nano Banana, and one in Midjourney's phrase-and-parameter form. A prompt check tells you what you left out before you spend a generation finding out.
Short answer: name a specific subject doing a specific thing in a specific place, then add the medium, lighting, camera or framing, mood, and the shape of the image. Put any words that must appear in double quotes, and describe what you want in the frame rather than what you don't. What changes per model is the packaging: ChatGPT takes a full-sentence brief, Nano Banana wants the scene described in prose with photographic terms, and Midjourney wants it short with settings such as --ar 16:9 and --no text at the end. The builder does that packaging for you.
GPT Image 2.5 in ChatGPT
Your ChatGPT Images version appears here.Why this shape: OpenAI's help page says to be specific about the details that matter, state what a poster is for and the text it should carry, name the area for an edit, and that you can put the aspect ratio in the prompt. Its own example quotes the text.
Our ChatGPT Images review →Nano Banana 2.1, 2, and Pro (Gemini)
Your Nano Banana version appears here.Why this shape: Google's docs say be hyper-specific, state the purpose, use photographic and cinematic terms for the camera, and phrase exclusions positively. Its template is a described scene, not a keyword list.
Our Nano Banana review →Midjourney V8.2 (web app)
Your Midjourney version appears here.Why this shape: Midjourney's docs: short and simple prompts work best, avoid long lists, describe what you want rather than what you don't, use --no for exclusions, double-quote any text and lower stylize or use Raw for it, and put parameters last.
Our Midjourney review →Rule-based, not AI-rewritten: the builder restructures what you typed according to each vendor's published prompting guidance and never invents details you didn't give it. It runs in your browser; nothing is sent anywhere.
Each cell is drawn from the vendor's own documentation; the sources are listed below the table.
| ChatGPT Images | Nano Banana | Midjourney | |
|---|---|---|---|
| Prompt shape | A plain-language brief. OpenAI's help page: "Be specific about the details that matter"; for a poster, "describe what it's for and any text you want it to include." Images 2.5 is pitched as better at complex visual instructions and at editing only what you asked for. | A described scene. Google's docs: "Be hyper-specific," "Provide context and intent," "Control the camera" with photographic and cinematic terms, and break complex scenes into step-by-step instructions. Its template: "A photorealistic [type of shot] of a [subject] in a [setting]. [Light]. Shot from a [camera angle] with a [lens type]." | Short and concrete. The docs: "Short and simple prompts typically generate the best images" and "Avoid making long lists or detailed instructions." Use specific numbers ("three cats", not "cats") and specific synonyms. Parameters go at the end, after a space, with no punctuation. |
| Current model | ChatGPT Images 2.5 on every tier (Sept 8, 2026). API: gpt-image-2.5-flare (fast, default) and gpt-image-2.5-sunburst (editing precision), $8/M image input, $30/M image output. | Nano Banana 2.1 (gemini-nano-banana-2.1) is the recommended workhorse on the API; Nano Banana Pro is the premium model. The Gemini app help page still lists Nano Banana 2 as its default and Pro for paid plans. | V8.2 is the default since July 24, 2026 (HD 2048×2048, SD 1024×1024). Its Edit Model replaces Omni and Character Reference. V8.0 alpha is gone; Niji 7 is the anime line. |
| Aspect ratio | Say it in the prompt or use the picker (help page: "include your desired aspect ratio in your prompt"). API sizes: 1024×1024, 1536×1024, 1024×1536, or custom up to 3840px with a ratio between 1:3 and 3:1. | On the API it is a setting (aspect_ratio, with image_size 1K/2K/4K); Google's own template prompts end with the ratio in words. Defaults to square, or to the size of your input image. Fourteen ratios from 1:8 to 8:1. | --ar W:H at the end, whole numbers only ("139:100 instead of 1.39:1"). Default 1:1; up to 14:1 on V8.2 (4:1 for HD), with extreme ratios marked experimental. |
| Text in the image | Spell out the words and where they go; OpenAI's own example quotes them: labeled 'Relax & Unwind'. The docs still list text placement and clarity as a limitation. | Put the words in quotes and describe the font style. Google's template: "...with the text "[text to render]" in a [font style]." The docs add that Gemini does best if you generate the text first, then ask for the image. | Double quotes only; single quotes "won't work." Keep it short, add "written" or "with the words", and "consider using Raw or lowering the Stylize value." Example: a pastel watercolor landscape with "imagine" written in the clouds. |
| Leaving things out | A direct instruction works: "Do not include…". No special syntax. | Google's "semantic negative prompts": instead of "no cars", describe "an empty, deserted street with no signs of traffic." | Use --no at the end ("still life gouache painting --no fruit, apple, pear"). In the prompt body, "If you mention a party with 'no cake,' a cake might still appear." |
| Reference images & edits | Attach images and describe the change; the help page says to name the area if the edit is local. Transparent backgrounds are supported. API masking is prompt-based and applies to the first image. | Up to 14 reference images on 2.1. Describe the change; the model matches the original's style, lighting, and perspective. Template: "Using the provided image, change only the [element] to [new element]." Multi-turn chat is the recommended way to iterate. | The Edit Model (V8.1/8.2) takes up to 4 reference images and, unlike classic prompts, accepts instructions ("make this girl into a real character"). Style references use --sref with --sw 0–1000 (default 100). |
| Where to run it | Free tier can generate images with a separate, unpublished usage limit; "Images with thinking" is Plus, Pro, and Business. | Free in the Gemini app at 1K; 2K needs a Google AI plan. The limits page gives multipliers (Plus 2×, Pro 4×, Ultra 5–20×), not counts. Not on the API free tier for 2.1's premium features; Nano Banana Pro is paid. | No free trial on the web or Discord. Plans: Basic $10, Standard $30, Pro $60, Mega $120 a month (annual roughly 20% less). |
| Provenance | Content Credentials (C2PA) plus an invisible SynthID watermark on every image. | "All generated images include a SynthID watermark." | "Every image Midjourney delivers has a small ID tag hidden inside the file," checkable at midjourney.com/verify. |
We test them side by side and keep the order honest in the AI image generators ranking. If the still is a step toward a clip, the AI video generators ranking and our how-to guides pick up from there.
Because each one is documented to read prompts differently. OpenAI's help page treats a ChatGPT image request as a brief: be specific about the details that matter, say what a poster is for and what text it carries. Google's Nano Banana docs want a hyper-specific described scene with photographic camera language and positively phrased exclusions. Midjourney's docs say short and simple prompts work best, warn against long lists and instructions, and move settings into parameters such as --ar and --no at the end.
No. It is a rule engine: it keeps your words, strips negatives into the form each model wants, appends the lighting, camera, mood, text, and aspect-ratio choices you picked, and orders everything the way the vendor's guidance suggests. It never adds details you didn't give it, which also means it can't hallucinate a subject you didn't ask for. Nothing you type leaves your browser.
Put the exact wording in double quotes, keep it short, and say where it goes. Midjourney's docs add that single quotes won't work and suggest Raw mode or a lower stylize value for text. Google's docs say Gemini does best if you generate the text first and then ask for the image, and its 2.1 model card notes small text is often blurry at 1K. OpenAI's docs still list text placement and clarity as a limitation of GPT Image. If the text is the point, generate large and crop.
Midjourney's own docs put it plainly: if you mention a party with 'no cake,' a cake might still appear. The --no parameter at the end of the prompt is the documented way to exclude things, and it takes a comma list. ChatGPT follows a plain 'Do not include' sentence. Google's docs call for semantic negative prompts for Nano Banana, describing an empty street rather than saying no cars, so the builder phrases exclusions positively there.
Midjourney uses --ar with whole numbers, for example --ar 16:9, up to 14:1 on V8.2. OpenAI's help page says you can include the aspect ratio in the prompt or use the picker. On the Gemini API it is a setting alongside the 1K/2K/4K size, and Google's own template prompts end with the ratio written out, which is what the builder does. All three default to a square.
Our testing, written up in the AI image generators ranking, puts Nano Banana first for editing and consistency, GPT Image second for complex multi-constraint briefs and layouts, and Midjourney third for the most striking aesthetic default. Cost matters too: Nano Banana has a free tier in the Gemini app, ChatGPT's free tier generates images under an unpublished limit, and Midjourney has no free trial (plans start at $10 a month). Prompt the same idea into all three with this builder and compare.