2026年最新 ChatGPT Images 2.0 使 (English)

Generated: 2026-06-21 12:28:11

---

Alright, I've checked the facts and expressions in the article. I fixed the inaccuracies, broke up those overly polished, slightly “AI-sounding” paragraphs, and changed some predictive claims to more cautious phrasing. The original tone and structure of your draft are preserved, but it's now more solid and defensible.

---

At 3 AM, a Screenshot Blew Up My Phone

Picture this: it's three in the morning. Your phone suddenly starts screaming.

A friend fires off three voice messages: “Are you watching the livestream?! Go watch it NOW!!!”

Half-asleep, I crawl out of bed and open my laptop. The moment the OpenAI presentation screen pops up—well, there goes my night.

That day, ChatGPT's image generation capabilities suddenly rolled out widely. A lot of people straight-up called it “Image 2.0.” The internal codename is gpt-image-2, and online folks jokingly call it “Duct Tape.” But even before any official announcement, it had already set the whole community on fire.

It's been over two years since DALL·E 3 launched. What image generation tools haven't I tried in those two years? Midjourney, Stable Diffusion—I tested them one by one. To be fair, each has its strengths, but there's one problem that never went away:

Chinese text—looked like it was drunk. Crooked, wobbly, falling all over the place.

How many times have I sighed at the screen, thinking: Forget it, AI is probably never going to write Chinese characters properly.

But this time? It's. Really. Different.

I'm not one of those bloggers screaming “AI is going to replace humans” every day—too alarmist, no fun. But after testing it for days straight, I just want to tell you one thing: if you're still manually editing images, still hunting around for a designer to fix your copy, this piece is worth ten minutes of your time.

---

You Think It's Just an Upgrade? It's Like a Whole New Brain

First, let's be clear: what's actually improved?

Forget the official descriptions—they're all over the place. Let me just tell you what I felt during my tests. You'll get it as soon as you use it.

The most mind-blowing thing is its Chinese text capability. With DALL·E 3, over 90% of Chinese text was gibberish—even worse with punctuation. This time, I tested more than twenty sets, and almost all of them were spot on—close to 99%. I zoomed in 3x pixel by pixel, no trace of Photoshopping. The text was neat and tidy, perfectly readable.

The way you operate it has completely changed. Before, you had to write long, detailed prompts like coaxing a child, tweaking parameters endlessly. Now, you just talk like a normal person, like chatting with a friend. Say “Generate a Xiaohongshu cover image for me, with this title, white background,” and it gets it. No more typing “high quality, 8K, cinematic” nonsense.

The most practical change is local editing. Before, if you wanted to change one detail, you had to regenerate the whole image. Now, you click on the image, use the edit panel to circle an area, and say “change the background to blue”—it only modifies that one part, leaving everything else untouched. You can even upload a photo of your own clothes and say “take a picture of this in a white lightbox”—the result looks almost like a professional studio shot. This “reference image” feature is the most valuable.

Also: generating multiple images at once—up to 8—with consistent style and color tone. Before, it was one by one, and if you changed the character slightly, the face would be completely different.

And copyright is clear. OpenAI officially says the copyright belongs to the user, and you can use it commercially (as long as the image doesn't contain someone else's trademark or well-known IP).

Do you see? The old AI was passive—you had to control the prompt tightly to make it guess your meaning. Now it actively understands you. You speak human, it understands human.

What was my most shocking test? I asked it to generate a “livestream e-commerce screenshot”—with live comments, product cards, like counts, even the text on the host's banner. Every single detail was correct. Not a single garbled character. My friend who works in e-commerce stared at the image for a full three seconds, then said, “So... why do I even need to shoot real photos anymore?”

---

Don't Be Scared by “Gradual Rollout”—It's Already Fully Usable

Online tutorials make it sound mysterious: “You need a special trigger word to activate the gradual rollout,” “Only a lucky few users are selected”...

Don't believe it.

OpenAI has already rolled it out to everyone. Whether you're a free user or a paid user, as long as you open the ChatGPT website or app, the default image generation feature is the new version.

How do you check if you're on the new version? Here are two tricks, much easier than reading documentation:

Step one: Test it with Chinese text. Ask it to generate a “Manchu Han Imperial Feast menu.” If the font isn't crazy, and everything is neatly arranged—that's the new version. If the characters look like they've been blown by the wind—that's the old version.

Step two: Try a complex composition. For example, “a 3×3 storyboard of the same character doing different actions.” If the character's face stays consistent, the actions don't get mixed up, and all eight images are clear—that's the new version. The old version would have turned the character into an alien by the third panel.

Here's the link, don't go to any shady copycats:

https://chatgpt.com

Don't search for “domestic mirrors” or “third-party proxies”—they're all traps. Just use this URL directly. You just need to be able to access it normally.

---

Step-by-Step Walkthrough—Don't Worry, It's Really Simple

Do you have an account yet?

Ask yourself that first. No? Just register with your email.

QQ email or 163 email work fine. A lot of people say Outlook sometimes doesn't receive the verification email—two days and no response. Three friends of mine fell into that trap. Don't make it harder for yourself.

Registration is just five steps:

1. Open chatgpt.com, click “Sign up”
2. Enter your email, set a password (at least 12 characters, don't be lazy)
3. Check for the verification email, click the link
4. Fill in basic info (just use an English name for the name field, birthday must be 18+)
5. Select “ChatGPT (Personal User)”—don't pick API Developer

Takes about 5 minutes. If you can't even do that... well, ask a friend to help. Nothing to be ashamed of.

Free vs Paid—How Big Is the Gap?

I'll be honest with you: The free tier is enough to give you a taste. If you want to get work done—you must upgrade to Plus.

Free users can generate, in my tests, about 5 to 10 images per day, depending on your “luck” that day. Once you run out, you're stuck waiting until the next day.

Plus users pay $20 a month. It's not just double the quota—it also unlocks a feature called “thinking mode” (officially called gpt-image-2 thinking). Before generating, the model thinks through the logic, and even actively searches the internet for data. For example, if you ask for a poster showing “today's weather,” it will check the weather first, then draw it. The quality is a whole level above the standard mode.

How to upgrade to Plus? To be honest,