What gpt-image-2 Means for Image Generation Inside Your AI Agent

In April 2026, OpenAI released gpt-image-2, its newest image generation model. If you use AgentBrush, this matters for one reason: it is the model AgentBrush generates with. You never call it yourself. Your agent calls AgentBrush, and AgentBrush calls gpt-image-2 for you.

Two things genuinely improved in this version, and they are the two things that used to embarrass image models: text inside images is now legible, and photorealism looks unstaged instead of plasticky. The best way to see what that buys you is to look at real prompts. Every example below is something you can ask your agent to make today. The image briefs are placeholders for now; we will generate the finished art in a later pass.

Text that is actually readable

Legible text was the headline upgrade, and it is the one that unlocks posters, thumbnails, ad creatives, and anything with a label on it.

A multilingual typography poster. Good for showing off the model's handling of scripts.

A professional editorial poster celebrating the world's writing systems. Swiss modernist style,
strong modular grid, generous negative space, refined typographic hierarchy. The greeting
"hello" set in eight scripts including Japanese, Arabic, Devanagari, Hangul, and Cyrillic, one
bold accent color against off-white. preset: flat_illustration · quality: high · 4:5 portrait

Generated with AgentBrush: flat_illustration preset, high quality, 4:5. The model set legible text in eight writing systems.

A launch poster with exact text. gpt-image-2 will render the strings you give it, so you can dictate the copy instead of hoping.

A Bauhaus-inspired launch poster: bold primary-color geometry, hard-edged shapes, off-register
print texture, confident asymmetry, a typography-led layout with a real ink-on-paper feel.

Poster text:
- Large title at the bottom: "SHIP IT"
- Headline at the top: "On-brand art, straight from your agent"
- Small footer: "agentbrush.dev"

Keep all visible text in English. preset: flat_illustration · quality: high · 4:5 portrait

Generated with AgentBrush: flat_illustration preset, high quality, 4:5. The exact headline, title, and footer text rendered as specified.

Photorealism that looks unstaged

The trick to convincing photos is asking for imperfection: real grain, uneven light, slightly off framing. Use the realistic preset and describe the camera, not just the subject.

A candid film photo.

A photorealistic candid 35mm film photograph of a flower-market stall at first light after rain.
Overcast morning, mixed neon signage and pale daylight, puddles mirroring the colors, a vendor
half-turned mid-motion, visible film grain, muted desaturated palette, natural imperfect framing,
the lived-in feeling of a documentary photograph. preset: realistic · quality: high · 3:2 landscape

Generated with AgentBrush, realistic preset, high quality. The film-grain and imperfect-framing cues sell the realism.

A casual phone snapshot. The "everyday photo" cues are doing the heavy lifting here, and the model will happily play along with something absurd.

A photorealistic iPhone snapshot of two raccoons sitting at an outdoor diner table in late
afternoon, taken casually by someone at the table. A shared basket of half-eaten fries, uneven
golden sunlight, relaxed slouched posture, slightly tilted framing, a hint of lens flare, the
natural imperfect realism of a real everyday phone photo. preset: realistic · quality: high

Generated with AgentBrush, realistic preset, high quality. The "everyday phone photo" cues do the heavy lifting.

Need the subject on a transparent background instead of in a scene? Generate it on a clean backdrop, then remove the background. That two-step workflow is in generating transparent-background PNGs.

One subject, every style

This is the part that earns a "whoa." Write the subject once, then run it through different presets to get a whole family of looks. Same idea, five finishes.

A friendly robot mascot holding a paintbrush, centered, simple background.
Run it once per preset: pixel_art, flat_illustration, isometric, logo, realistic.

The same robot mascot shown across five styles. Generated with AgentBrush.

For a style the presets do not cover, reach for custom and describe it yourself. gpt-image-2 has deep art-history knowledge, so specific movements work well.

A traditional Japanese ukiyo-e woodblock print of a heron rising over reed marshes at dusk.
Layered flat color planes, delicate carved linework, soft gradient sky, aged washi-paper texture,
subtle off-register ink, a muted indigo and ochre palette. preset: custom · quality: high · landscape 3:1

Generated with AgentBrush, custom preset, high quality, 3:1. Custom handles styles the presets do not cover.

Cinematic game key art. When you want a dramatic, near-photoreal scene for a hero shot or a store page, describe it like a film still. The realistic preset carries the lighting, depth, and detail.

Cinematic game key art: two pirate crews about to clash as their ships close in heavy seas, a
weathered galleon flying the Jolly Roger, cutlasses drawn, a sword-arm thrust into the
foreground, bright hazy daylight over a rugged coastline, photoreal detail, epic wide
composition. preset: realistic · quality: high · landscape 3:2

Game key art generated with AgentBrush, realistic preset, high quality. Great for a hero shot or store page.

A note on shape: AgentBrush supports custom sizes up to a 3:1 ratio, so the 4:5 portraits above and this wide 3:1 banner both come straight from the size you ask for. (More in tokens, quality, and cost.)

Start from an image you already have

gpt-image-2 is strong at working from an input image, which is how you get consistency and edits rather than a fresh roll of the dice each time. In AgentBrush this is the reference-image feature.

From one portrait, many outputs. Register a photo as a reference, then ask for anything. The same person carries into every result. Here is one studio headshot, used two ways.

First, a data-dense color guide:

Using this portrait, create a diagram-first personal color analysis. Show which clothing colors
suit the subject through labeled swatches and side-by-side comparisons. Keep text minimal.
preset: custom · with the portrait as a reference image

Then the same face as a flat avatar:

Turn this portrait into a clean flat-illustration avatar. Keep the same face, hair, and skin tone.
preset: flat_illustration · with the portrait as a reference image

One reference photo, two very different outputs, the same person in both. Generated with AgentBrush, high quality.

The same product, a new scene. Pass a product photo as a reference and AgentBrush keeps it identical while you change everything around it.

Place this exact travel mug on a sunlit wooden cafe table beside a laptop and a small plant,
soft morning light, lifestyle product photo. Keep the mug identical.
preset: realistic · with the mug as a reference image

Identical mug, brand-new lifestyle shot. The reference holds the color, shape, and cork base steady.

Want to repaint just one region instead of regenerating the whole image? That is the mask editor, covered in the mask editor walkthrough.

From prompt to project

Every prompt above is one line your agent sends to AgentBrush from inside Claude or Cursor. AgentBrush compiles it into a well-ordered gpt-image-2 call, picks the size, tracks the token cost, and saves the finished file into your project. No API key in your shell, no upload glue, no tab-switching. The setup takes about two minutes: see installing AgentBrush.

FAQ

Is gpt-image-2 better than DALL·E 3? For following prompts, rendering text, and editing, yes. It is a clear step up, and the DALL·E 3 API is being retired, so for new work the choice is mostly made for you.

Do I need my own OpenAI API key? No. Your agent calls AgentBrush, and the generation runs on gpt-image-2 behind a single subscription.

Can it generate a transparent background? Not in one step. Generate on a clean background, then remove it. AgentBrush does the removal locally and for free.

How do I control the aspect ratio? Ask for it in the prompt and AgentBrush maps it to the image size. Standard squares and portraits are built in, and custom sizes go up to a 3:1 ratio.

Want to make any of these from your editor? Connect AgentBrush to your agent and paste one of the prompts above. The first image lands in your project in seconds.