GPT Image

byOpenAI

Filmic light, correctly spelled type and surgical regional edits, from a first sentence to a finished hero frame

GPT Image

How GPT Image works

Write naturally, attach references or a mask only when you need them, then choose a tier. Draft cheap and quick, then re-run the keeper at higher quality.

Write the scene in full sentences

Write the scene in full sentences

Describe subject, light, materials and any exact text in quotes, using plain natural language instead of tag lists.

Attach references or a mask

Attach references or a mask

Drop in one or many reference images to blend, or add an alpha mask so only the marked region gets rewritten.

Pick a tier and render

Pick a tier and render

Choose the model, quality level and resolution up to 3840x2160, toggle transparency for text-only prompts, and run up to four at once.

What GPT Image is good at

Built around three jobs most image tools fumble: legible text inside the artwork, edits that touch only the marked region, and blending several references without losing who or what is in them.

Fast and flagship tiers, up to 4K

Fast and flagship tiers, up to 4K

Flare is the quick daily driver for concept drafts at up to 4K; Sunburst takes over for final hero assets and the xhigh and max quality tiers. Quality sets sampling effort, not pixel size, so draft low and finish high.

Native alpha transparency

Native alpha transparency

Generate an isolated subject straight onto a clean alpha channel for stickers, UI icons and product mockups, with no cutout step. Transparency works in text-only mode, so leave reference images off.

Mask-based regional inpainting

Mask-based regional inpainting

Upload a base image plus a same-size mask with an alpha channel; transparent zones are the only pixels rewritten. Surrounding light, grain and layout stay put, and 1.5 remains a dependable pick when strict preservation matters.

Multi-image fusion with identity lock

Multi-image fusion with identity lock

Feed several references and tell the prompt how to combine them: palette from one, subject from another. Sunburst is the tier built for complex reference fusion and multi-turn edits that keep a character recognisable.

Made with GPT Image

A spread of campaign, fashion, product, film and sticker work, with typography, transparent cutouts and filmic lighting in the same gallery.

Editorial lookbook with sharp colour contrast

Editorial lookbook with sharp colour contrast

Product hero on natural substrate

Product hero on natural substrate

Cinematic music-video frame

Cinematic music-video frame

Graphic poster with legible multilingual type

Graphic poster with legible multilingual type

Transparent sticker and UI asset render

Transparent sticker and UI asset render

What people build with GPT Image

Used by teams that need real words in the picture, controlled edits, and clean assets: campaign designers, lookbook creators, product-asset builders, retouchers and spatial designers.

Multilingual campaign poster teams

01

Put exact headlines in Latin and non-Latin scripts directly inside the artwork and re-run the same layout for each market. Type stays crisp and kerned, so localized hero graphics don't need a manual typesetting pass.

Lookbook and character-consistency designers

02

Blend model, garment and location references into one frame and keep the same identity across iterations. Reach for Sunburst when a fusion has to hold together under close inspection across several rounds of edits.

E-commerce and sticker asset builders

03

Produce isolated product shots, icons and stickers on clean transparency at batch scale. The budget Mini tier suits high-volume thumbnails; step up when the hero image needs more texture.

Photo editors doing regional retouches

04

Mask one area, describe the change, and leave the rest of the frame untouched. Swap a prop, relight a surface or replace a sign while the surrounding grain and lens depth hold steady.

Architects and set designers

05

Turn a spatial brief into cutaways, interiors and wide panoramas at 2K or 4K, with consistent perspective and believable light. Pair high quality with large formats for crisp structure.

GPT Image specs

Plans and credits
Versions
GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, GPT Image 2, GPT Image 1.5, GPT Image 1, GPT Image 1 Mini
Inputs
Prompt (text), Images (image), Mask (image)
Output
image
Aspect ratios
2560x1440 (2K), 3840x2160 (4K), 1024x1024, 1536x1024, 1024x1536
Cost
7–552 credits per run

GPT Image questions

Flare is the faster tier and Sunburst is the higher-fidelity one. Flare suits rapid concept drafting and daily work at up to 4K with low latency, but gives slightly less nuanced detail and reference blending. Sunburst is for final hero assets, complex multi-image fusion and the xhigh and max quality tiers, at the cost of higher latency and higher generation cost.

Use GPT Image 1 Mini, the most cost-effective option for batch generation and thumbnail pipelines. It trades away some spatial precision on dense scenes and renders simpler textures, so move up to Flare or Sunburst when the image needs more detail.

Pick 2.5 for 4K output and the extra quality tiers, and 1.5 when you want a proven workhorse for text-to-image and inpainting that preserves the surrounding layout. GPT Image 1.5 lacks the 2.5 series' expanded quality tiers and native 4K modes.

Turn on the Transparent toggle and generate from a text prompt alone. Transparency is only supported in pure text-to-image mode, so it won't work if you attach reference images. Describe an isolated subject for the cleanest edges.

Save the mask as an image with an alpha channel, at exactly the same size and format as your base image. Transparent areas mark the zone to be edited. A black-and-white RGB mask without alpha can cause processing errors.

It is not built for sub-second generation, uncensored content, or replicating niche styles that depend on LoRA fine-tunes. Known artifacts include symmetry warping in intricate mechanical objects, limb fusion in crowded group scenes, and garbled text in long background signage.

Try GPT Image on Fuser

Filmic light, correctly spelled type and surgical regional edits, from a first sentence to a finished hero frame

Explore all recipes
All articles