Most ChatGPT image attempts fail the same way: a 30-word adjective salad (“beautiful cinematic detailed mystical glowing fantastical 4k”) produces a generic picture, and the user gives up after a dozen rolls. People who get usable assets do the opposite — they write short, structured prompts (subject + style + lighting + camera) and change one variable at a time. As of June 2026, ChatGPT runs the new ChatGPT Images 2.0 model (gpt-image-2, shipped April 21, 2026), which plans before it draws, edits a selected region without redrawing the whole frame, and finally renders readable text. This is the workflow that takes advantage of all three.
TL;DR
- ChatGPT’s image model is ChatGPT Images 2.0 (
gpt-image-2), live since April 21, 2026. It replaced GPT Image 1.5; DALL-E 3 was deprecated May 12, 2026. - Write a 7-line structured brief (subject, action, style, lighting, camera, mood, avoid). Generate once, find the single worst miss, change only that one line, repeat. Three rounds, not thirty.
- Use the selection (mask) tool for local fixes — highlight a region and describe the change — instead of regenerating the full image.
- “Thinking” mode (the model self-reviews and can search the web for references) and 2K output require a Plus, Pro, or Business plan. Free is roughly 2-3 images per 24 hours.
- Text rendering is now genuinely good, including Chinese, Japanese, Korean, and Hindi. The old “add text in post” rule no longer applies for short labels.
What ChatGPT Images 2.0 actually is (June 2026)
ChatGPT’s built-in image generation is native to the chat model — you describe an image, refine it in plain language, and edit it in the same thread. The current model is gpt-image-2. Unlike the older diffusion pipeline, it runs a short reasoning loop before drawing: it plans the composition, can look up reference material on the web, generates candidates, and checks the result against your prompt. That is why a vague brief now produces a competent-but-bland image instead of an outright broken one — and why a precise brief gets you most of the way in one or two rounds.
| Capability | As of June 2026 |
|---|---|
| Model | ChatGPT Images 2.0 (gpt-image-2), launched Apr 21, 2026 |
| ”Thinking” mode (self-review + web reference search) | Plus, Pro, Business only |
| Resolution | Up to 2K (2048×2048) in app; 4K is a beta API flag |
| Aspect ratios | From 3:1 (wide) to 1:3 (tall) |
| Multi-image consistency | Up to ~8 coherent images of the same subject from one prompt |
| Reference images for editing | Up to 10 uploaded images combined in one generation |
| Text rendering | Strong, including Chinese, Japanese, Korean, Hindi |
| Free plan | ~2-3 images per rolling 24 hours |
| Plus ($20/mo) | ~50 image prompts per rolling 3-hour window |
| Pro ($200/mo) | Unlimited, faster image creation (subject to abuse guardrails) |
Use it for: blog cover art, social posts, landing-page hero variations, conceptual diagrams, simple product mockups, and mood boards. Skip it for: real product photography, brand-critical hero art that must stay pixel-consistent across 50 assets, and designs needing exact typographic layout. For deep stylistic control, Midjourney still wins; for precision layout, hand-design.
Before you generate anything
- Write a one-sentence brief: who is in the image, what they are doing, the mood, and where it will be used. Without it, every prompt drifts generic.
- Grab one reference image — even your own past work — so you can say “like this, but with X different.” Underused, and the single biggest quality lever.
- Pick the aspect ratio first. 16:9 for blog headers, 1:1 for social, 9:16 for stories. The model supports 3:1 to 1:3, so state it explicitly. Baking composition at the wrong ratio wastes every roll.
The 7-line prompt structure
Subject: senior software engineer, mid-30s, looking thoughtful
Action: writing in a notebook at a wooden desk
Style: editorial illustration, muted color palette
Lighting: soft morning window light, warm tones
Camera: medium shot, slight angle from the left
Mood: contemplative, focused
Avoid: extra fingers, multiple people, dark backgroundThis structure produces predictable results because it forces a decision on every axis the model would otherwise improvise. Skip a line and gpt-image-2 fills the gap, usually in a direction you did not want. Camera and lighting matter more than the subject for how “professional” an image reads — most weak outputs are missing exactly those two lines.
Step by step
- Write the brief in the 7-line structure above. Fill every line.
- Generate once. Read what you got and identify the ONE thing most wrong (composition, lighting, palette, or pose).
- Iterate by changing only that one line: “Same composition, warmer lighting.” “Same lighting, pull back to a wide shot.” Changing three things at once means you will not know which change helped.
- For local fixes, use the selection tool: highlight the region (drag the size slider to adjust the brush), then describe the change in chat — “replace the background with a soft blue gradient” or “change the laptop to a paper notebook.” This edits only the masked area and keeps the rest intact. Do not regenerate the full frame for a small fix.
- For a consistent character across a series, upload round 1’s image as a reference in round 2 and keep the style lines identical, changing only the action or setting. The model can hold facial features, clothing, and palette across up to ~8 images.
- Save the final prompt and image together. A simple convention like
topic_style_lighting.pngnext to the prompt text makes the asset reusable next month.
Stop at three iterations. If it is not close by then, the brief is wrong, not the prompt — rewrite the brief instead of rolling again.
Before you publish
- Does the image match its destination? An anime portrait does not work as a B2B blog header, no matter how good it looks.
- Scan for hallucinated artifacts — extra fingers, melted shapes, the wrong number of windows. The reasoning loop reduced these but did not eliminate them.
- Confirm the aspect ratio matches where it will live. A 1:1 composition cropped to 16:9 loses the framing you built.
- For brand work: would this sit next to last week’s image without looking like a different artist made it?
Make the workflow repeatable
- Keep a
prompts.mdlibrary, sectioned by use case (blog headers, social, mockups). Each entry: brief, prompt, result image, and what you learned. - For recurring needs (a weekly newsletter header), pin the working prompt and change only the topic noun each week.
- Build a Custom GPT for your visual brand: put your style words, color palette, and “avoid” list in the Instructions. Every prompt then starts from your brand baseline instead of zero.
Total time per usable image with this loop: 5-10 minutes, versus the open-ended roulette of unstructured prompting.
Common mistakes
- Stuffing 10+ adjectives into one prompt — they fight each other and produce mush.
- Omitting the camera and lighting lines, which decide more than the subject does.
- Regenerating the whole image for a small fix instead of using the selection tool. It wastes rolls and breaks consistency.
- Changing three variables between iterations, then not knowing which one worked.
- Ignoring aspect ratio until the end. A composition baked at 1:1 will not crop cleanly to 16:9.
- Assuming text is still broken. On
gpt-image-2, short labels and headlines render reliably; only paragraph-length text is still risky.
FAQ
- Why do my images look generic?: Almost always a missing camera or lighting line. Add both, even if you have to look up a couple of cinematography terms (“medium shot,” “soft window light,” “Rembrandt lighting”).
- How do I get the same character across multiple images?: Upload round 1’s image as a reference in round 2 and keep the style lines identical, changing only the action or setting.
gpt-image-2holds facial features, clothing, and palette across up to ~8 images from one prompt. - Can it render text now?: Yes — this is the biggest change. ChatGPT Images 2.0 renders short text reliably, including Chinese, Japanese, Korean, and Hindi. For full paragraphs, still add the copy in a design tool afterward.
- Why does it refuse certain images?: Content policy. Requests involving real or famous people, copyrighted characters, and violent or explicit content are declined.
- How many images can I make per day?: As of June 2026, Free is roughly 2-3 per rolling 24 hours, Plus around 50 prompts per rolling 3-hour window, and Pro is effectively unlimited and faster. The “thinking” mode and 2K resolution need Plus or higher.
- How is this different from Midjourney?: ChatGPT is easier to iterate — you describe edits in natural language and mask regions directly — and now matches or beats Midjourney on text and layout. Midjourney still wins on stylistic depth and painterly finish. Many people use both.