GPT Image·by OpenAIText to image
GPT Image 2.6 Flare Text-to-Image
Image generation from natural-language prompts with OpenAI GPT Image 2.6. 14 sizes up to 4K (3840×2160), five quality tiers including xhigh and max, true transparent backgrounds, and a free complexity setting for how much the model plans before rendering. Prompts up to 32,000 characters. Priced live from quality, size and prompt length.
Parameters
| Parameter | Type | Default | Range or options |
|---|---|---|---|
Output format output_formatOutput image format. PNG keeps transparency; JPEG files are smaller. | select | png | png, jpeg |
Background backgroundtransparent produces a real alpha channel (needs PNG output). This setting is how OpenAI makes a transparent background on GPT Image 2.6 — describe one isolated subject with nothing behind it. | select | auto | auto, opaque, transparent |
Complexity complexityHow much the model plans the image before rendering, separately from quality. high (the model default) helps with dense text, complex layouts and many objects; low suits simple images and focused edits, and is faster. No extra charge. | select | high | low, medium, high |
Moderation moderationContent-moderation strictness on the provider side. low applies less restrictive filtering. | select | low | low, auto |
Quality qualityRendering quality tier; xhigh and max sit above high. At 1024×1024: low ≈0.6×, high ≈3.2×, xhigh ≈5.4×, max ≈11.9× the medium price. | select | medium | low, medium, high, xhigh, max |
Size sizeOutput size in pixels — the only control over dimensions and aspect ratio. Sizes above 2560×1440 are experimental upstream. Larger sizes cost more. | select | 1024x1024 | 1024×1024 · 1:1 (1024x1024), 1024×768 · 4:3 (1024x768), 768×1024 · 3:4 (768x1024), 1024×1280 · 4:5 (1024x1280), 1536×1024 · 3:2 (1536x1024), 1024×1536 · 2:3 (1024x1536), 2048×2048 · 1:1 (2048x2048), 2048×1152 · 16:9 (2048x1152), 1152×2048 · 9:16 (1152x2048), 2560×1088 · 21:9 (2560x1088), 1088×2560 · 9:21 (1088x2560), 2880×2160 · 4:3 (2880x2160), 2160×2880 · 3:4 (2160x2880), 3840×2160 · 16:9 (4K) (3840x2160), 2160×3840 · 9:16 (4K) (2160x3840) |
Inputs
No reference files: this model works from the prompt alone.
LUVI Prompt Engine rules for GPT Image
TARGET MODEL: GPT Image 2.6 (OpenAI — Flare and Sunburst read the same dialect). The prompt reaches the model exactly as written. Prose, reference images and the app's settings (size, quality, complexity, background) are the whole control surface — there is no seed, no negative_prompt field, no weighting and no special token syntax.
Rules that must hold
- Keep the user's subject, intent and every quoted string exactly, and every style or medium word they gave, in English when you translate ("fotoğraf gerçekçi" → "Photorealistic photograph.", "yağlı boya" → "oil painting", "3D render", "watercolour") — dropping one changes the look of the result. Reshape; do not invent. Never add props, scene elements, people, wardrobe, architecture, materials, weather, time of day, light, lenses, styles or orientation words ("vertical", "wide") the user did not give. A short brief stays short: "a red vespa in a narrow Istanbul street" must NOT gain cobblestones, shuttered windows or afternoon light — and do not restate the brief in a second sentence.
- Every format works on this model (short prompts, prose, tags, JSON-like, labeled sections). If you restructure a JSON-like or tag prompt, carry over every value — no field may be dropped or reinterpreted.
- Name the deliverable and its use in the opening words when the user gave them: "a vertical retail poster for a coffee roastery", "a product photograph for a web shop", "an infographic for high-school students".
- In-image text: every visible string in double quotes, exactly as it must read, with its position and typography in plain words; say how many times it appears when repetition would hurt ("exactly once"). A string the user put in quotes — single quotes, typographic quotes or double quotes — is written in double quotes in your output, letter for letter: 'Bayramınız Kutlu Olsun' becomes "Bayramınız Kutlu Olsun". Never turn quoted copy into plain words. Keep non-English characters exact (Turkish İ ı ş ğ ç ö ü, ₺). Spell an unusual brand name letter by letter only if the user did.
- A name the user gives for a product, shop or brand is printed by this model whenever the image has room for it, and it invents label copy around it (live: "a coffee bag from a roastery called Kuzey Kavurma" came back with "Kuzey Kavurma" plus an invented origin, tasting notes and weight). So when the image shows the thing that carries the name — the brand's packaging, label, bottle, sign, storefront, logo, menu or poster — always add one sentence that quotes the name and says where it is printed, then end with "No other text.": "… on a wooden shelf. The bag's label reads "Kuzey Kavurma". No other text." (live: only the name, no invented copy, 2/2). Copy the user gave beside the name (a slogan, a price, a line) is quoted in the same sentence with its place, and the name stays in it: "The can reads "<name>" and, below it, "<slogan>"." A name that only describes a person or place in the scene ("a barista who works at a café called Moka") gets no such sentence.
- Exclusions are short plain sentences at the end: "No other text." · "No watermark, no logos." Keep every exclusion the user wrote. When the image carries quoted text, end with "No other text." (OpenAI's guide asks for no extra text around required copy; live, it removed invented lettering).
- When the user says the image should carry no text (in any words: "no text", "without text", "yazısız", "yazı olmasın") and the scene invites lettering — a town, street, harbor, market, shopfront, café, station or vehicle — write that exclusion as "No text, signs or lettering anywhere." (live: a plain street scene came back with a chalkboard menu and a flag; with this sentence, none). When the user said nothing about text, add no text exclusion: signs in a street are part of the scene. Add no other exclusion of your own.
- Transparent background is a setting (background=transparent, PNG). When the user wants a cutout, describe ONE isolated subject with nothing behind it — no backdrop, floor, scene or colour — and keep any "transparent background" words the user wrote.
- Aspect ratio, resolution, quality and complexity are settings. Do not write ratio numbers ("16:9"), resolutions ("4K", "8k") or "low/high complexity" into the prompt; describe the composition in words instead ("a wide landscape", "a tall vertical poster with empty space at the top for the headline").
What works best
- For a long brief use the order scene → subject → details → constraints; labeled sections ("Scene:", "Subject:", "Details:", "Text:", "Constraints:") are fine.
- Describe what is visible: materials, light, colour, medium, framing, texture. Ask for "photorealistic" or "real photograph" only when realism is the user's goal; camera terms ("35mm film photograph, eye level, 50mm lens") are look cues, not physics.
- Replace mood words and quality boosters ("cinematic", "stunning", "8k, masterpiece, best quality") with the scale, atmosphere, colour and concrete detail the user already gave — never with new content.
- People: body framing, relative scale, gaze and what the hands do ("full body visible, feet included", "looking down at the open book"), when the user described the person.
- Charts, slides, diagrams: keep the real numbers and labels in the prompt.
From the LUVI Prompt Engine rules for this family — the same rules the workspace rewriter and LUVI's MCP connector apply.