ChatGPT’s image generation runs differently from Midjourney: there are no --ar flags, the model reasons about your request before it renders, and it can keep a character consistent across a whole set in one conversation. This walks you from “type the prompt” to a publishable image, then to a style-consistent series for social channels.
TL;DR
- As of June 2026, ChatGPT image generation runs on ChatGPT Images 2.0 (model
gpt-image-2), which OpenAI shipped April 21, 2026 to replace GPT Image 1.5. - Just type your request in the normal chat box. Words like “generate”, “draw”, “image”, or “create a picture of” trigger it; you don’t open a separate tool.
- Free users get roughly 2-3 images per day on the base model. Plus ($20/mo) unlocks Thinking mode (reasoning, character consistency across up to 8 images) at roughly 50 images per rolling 3-hour window.
- Aspect ratios run 3:1 to 1:3 at up to 2K resolution, set in plain language (“make it 4:5”, “vertical 9:16”). No Midjourney parameters.
- The real edge over Midjourney: it remembers the previous image in the same chat, so “same character, new scene” actually holds.
What ChatGPT Images 2.0 changed
The April 2026 update is worth knowing because the workflow advice below depends on it. Three things are new versus the older DALL·E behavior:
| Capability | Before (GPT Image 1.5) | Now (ChatGPT Images 2.0) |
|---|---|---|
| Images per prompt | 1 | Up to 8, with character/object continuity |
| Aspect ratios | Limited presets | 3:1 to 1:3, recomposed (not cropped) |
| Reasoning step | None | Plans layout and text placement before rendering (Plus+ Thinking) |
| In-image text | Often garbled | Readable typography, including Chinese, Japanese, Korean |
| Reference editing | Coarse | Upload an image, select a region, describe the change |
The reasoning step matters most: on Plus and above, the model interprets spatial relationships and text placement first, so infographics, menus, and labeled diagrams come out legible instead of as gibberish.
Where it lives and which tier you need
There is no separate image panel. You generate from the regular chat box.
| Tier | Price (June 2026) | Image access | Rough limit |
|---|---|---|---|
| Free | $0 | Base model only | ~2-3 images/day |
| Go | $8/mo | Base model | Higher than Free |
| Plus | $20/mo | Base + Thinking mode | ~50 / 3-hr window |
| Pro | $200/mo | Base + Thinking + ImageGen Pro | Highest |
OpenAI does not publish exact image caps, and they shift; treat these as community-observed ranges as of June 2026. The practical line: if you want consistent characters across a set, you need Plus or above so Thinking mode is available.
The minimum prompt that works
You don’t need Midjourney parameters, but structure still beats a vague sentence. Order the prompt like this:
Subject + Style + Lens / POV + Lighting + Mood + AspectExample:
A 30-year-old man drinking coffee by a cafe window, realistic photography style, 50mm f/1.8 lens, natural window light, warm and quiet mood, 4:5 aspect.
The two parts most people skip are lens and lighting direction, and those are exactly what separate a real-looking photo from “AI stock”. A named focal length (50mm, 85mm) plus a light source (“side-light from a window”, “overcast soft light”) does more than ten mood adjectives.
For aspect ratio, just say it in words. ChatGPT recomposes the frame for the new ratio rather than cropping, so the same idea ships cleanly as 1:1, 9:16, and 16:9 for different placements.
How to keep a series consistent
This is ChatGPT’s killer feature: it remembers the previous image in the same chat. To use it:
- Generate image 1, then name exactly what to preserve: “keep this lighting and color grade, same character and outfit”.
- For each new image, anchor to the prior one: “Based on the previous image, same person, now in a kitchen at night.”
- After 3-4 iterations, freeze the style into a fixed block of text and only change the subject line. Paste the same block every time.
- On Plus or above, ask for the whole set at once: “Generate 4 images of this character in different scenes, consistent face and outfit.” Thinking mode holds continuity across up to 8 images in a single request.
Two things that break consistency: starting a new conversation (style resets) and editing the frozen style block mid-series. Keep one chat, keep the block stable.
Editing an existing image
Upload a photo and you can either select a region and describe the change, or describe a broader edit in conversation. Plain-language follow-ups steer the result without starting over: “remove the cup”, “change the jacket to navy”, “fix the hands”. Region selection is not pixel-perfect, so for tight masks, expect a couple of passes.
What it’s good for
- Xiaohongshu / blog covers, banners, and thumbnail variants
- Personal avatars and virtual-character exploration
- Product mock-ups and scene shots for design conversations
- App / website hero references and storyboards
- Infographics and labeled diagrams (the readable-text upgrade makes this newly viable)
What it’s not good for
- Vector, scalable, or fully editable final deliverables (logos, icons) — it outputs raster pixels, not editable paths
- Strict commercial-rights assets where you need clear licensing for trademarks
- Exact text on a tight deadline — typography is far better now but still proofread it
On rights: under OpenAI’s current usage terms, you own the images you create and can use them commercially, but don’t lean on AI for logos, trademarks, or anything needing strict IP clearance. Check the latest terms before a commercial launch.
FAQ
Where is ChatGPT image generation in the interface? It lives in the regular chat box for Free, Plus, and Pro users — there’s no separate panel. It triggers on words like “generate”, “draw”, “image”, or “create a picture of”. Free users are limited to roughly 2-3 images per day.
Do I need Midjourney-style parameters like --ar?
No. ChatGPT doesn’t parse --ar, --v, or --sref. Use natural language for ratio (“4:5”, “vertical 9:16”). It supports 3:1 to 1:3 at up to 2K resolution, and it recomposes the frame for the new ratio instead of cropping.
How do I keep a 4-image series visually consistent? Generate everything in one conversation and reference the previous image explicitly (“same character, same outfit, now in a kitchen”). On Plus or above, Thinking mode can produce up to 8 images with continuity in a single request. Starting a new chat resets the style.
Why does my image look generic or “AI-stock”?
The prompt is missing specifics. Concrete lens and lighting direction beat generic adjectives every time. Add a 50mm f/1.8, side-light from a window, and a named mood before re-running.
What’s the difference between Free and Plus for images? Free runs the base model only (~2-3 images/day). Plus ($20/mo) adds Thinking mode — the reasoning step that plans layout and text, plus character consistency across up to 8 images — at roughly 50 images per 3-hour window. Pro adds an ImageGen Pro layer on top.
Related
Tags: #ChatGPT #Image generation #Prompt