
Variant families explained
GPT Image 2 is the neutral route, the VIP route adds 2K output, and the 2.5 and 2.6 Flare and Sunburst variants trade some neutrality for a stronger look.
1 free try-on, 1 free video and 1 free image every day — no sign-up neededSee what's free
OpenAI’s image model, in the studio: GPT Image 2 and 2 VIP, plus the 2.5 and 2.6 Flare and Sunburst variants for a more stylised take. 1K or 2K output, up to ten reference photos, and the best of the three at following a long, specific instruction.
Log in to generate and keep your images.
Your generated images appear here.
Describe an idea, pick a model and press Create. Four renders of one prompt usually land one image worth keeping.
Every image model lives on one page too, with reference uploads, batches and the full resolution list. AI Image Generator
GPT Image 2 answers a written brief, not a keyword list. Say what is in frame, where it sits and what the style is, and add up to ten references when the layout has to match something that already exists.
"A bottle at the lower left, a linen cloth behind it, morning light from the right." Spatial language is what this model is best at, and it is the reason people pick it over the diffusion models.
If the composition is right and one object is wrong, edit instead of re-prompting. Each regeneration costs an image, and this model is not the one the free daily tier covers.
OpenAI’s image model is an instruction follower: the more precisely you place things, the more of them come back where you asked.
Up to 10 references.Attach a brand sheet, a packaging photo or a mood board, then describe the differences in words. GPT Image copies proportion and typography from a reference better than it invents them from a description.
Object. Position. Light. Style.Four clauses, in that order. It sounds mechanical and that is the point: GPT Image follows a structured sentence literally and then stops, which is exactly what you want from a layout.
Layouts, packaging and anything with words in it — the jobs where a picture that is ninety percent right is useless.

A label wrapped around a bottle has to sit on the curve and stay legible. GPT Image gets the wrap right more often than the diffusion models, and 2K is enough for a shop render.
Notes from sellers, designers and writers who needed one picture, not a photoshoot.
“I keep one reference photo of each product and describe a new background. Ten listing images in an afternoon, and the free image is what made me try it at all.”
“Four renders of one prompt is the useful part. I post the one that still reads on a phone, and the other three cost nothing to look at.”
“The reference photo is what makes product shots usable. I upload the real item, describe a new surface, and the render keeps the shape, the label and the proportions.”
“I sketch a direction with prompts, then draw the real thing. Twelve cheap thumbnails beat staring at a blank page.”
“The button told me the price before I clicked, and the one render that failed came back as credits. That is all I ask from a paid tool.”
“Image to image keeps a character on model. I upload the sheet I drew, describe a new pose, and the style survives.”
“I keep one reference photo of each product and describe a new background. Ten listing images in an afternoon, and the free image is what made me try it at all.”
“Four renders of one prompt is the useful part. I post the one that still reads on a phone, and the other three cost nothing to look at.”
“The reference photo is what makes product shots usable. I upload the real item, describe a new surface, and the render keeps the shape, the label and the proportions.”
“I sketch a direction with prompts, then draw the real thing. Twelve cheap thumbnails beat staring at a blank page.”
“The button told me the price before I clicked, and the one render that failed came back as credits. That is all I ask from a paid tool.”
“Image to image keeps a character on model. I upload the sheet I drew, describe a new pose, and the style survives.”
Instruction following is a skill you can also use badly: four rules.

GPT Image 2 is the neutral route, the VIP route adds 2K output, and the 2.5 and 2.6 Flare and Sunburst variants trade some neutrality for a stronger look.

Use them for what words describe badly: a logo, a typeface, a packaging finish. Not for ten mood images that pull the style five different ways.

Text renders more reliably here than elsewhere, but at 1K a one letter error is invisible until print. Render the final at 2K and zoom in before it leaves the studio.

A batch costs per image and repeats the same prompt. Vary the prompt between renders instead of hoping the seed changes the composition, because it will not.
Which variant to use, what the references do, and where instruction following beats diffusion.
GPT Image 2 and GPT Image 2 VIP, plus four stylistic variants: 2.5 Flare, 2.5 Sunburst, 2.6 Flare and 2.6 Sunburst. The VIP route adds 2K output; the Flare and Sunburst variants trade some neutrality for a stronger look.
The free daily image is reserved for the cheapest Nano Banana tier, so GPT Image renders are paid from credits. The price per image appears before you generate, a batch is charged per image, and a free account is all you need to top up.
Instruction following. Layouts with text, several objects placed deliberately, and edits that only touch the part you describe are where it beats the diffusion-first models. Write the brief like a brief — what is in frame, where it sits, what the style is — and skip the keyword pile.