English | Chinese Original

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Date: 2026-04-21

Prolonged dry cargo warning: this document contains a large number of complete messages that can be copied directly, without loss

Hello, I'm cxuanai.

My friends' circle and Twitter were swiped by something called GPT-image-2 for the last two days. I didn't care much at first, after all, I've seen a lot of this stuff in AI. Midjourney, Stable Diffusion, Nano Banana, who didn't just come out with a little bit of a blowout, and it's either flipping or flipping.

But this is a little different.

It started in early April 2026. At that point, three anonymous models appeared on the LMArena assessment platform, code-named maskingtape-alpha, packingtape-alpha, gaffertape-alpha, with names that looked like a joke -- masking tape was a strip of paper, packing tape was a tape, gaffer tape was a tape tape, all of which was tape. The three models disappeared a few hours later, but the samples that came out of the web had already exploded.

After a deep-seated attack, the API returned metadata with GPT-image-2 writing. OpenAI has no official declaration, but DALL-E 2 and DALL-E3 have decided to officially suspend their uniforms on 12 May this year, apparently to give way to the new king.

So what the hell is this? In human terms, the next generation of raw image generation models, code Spud, launched by OpenAI. Its greatest feature is that ** it's in Tulipi and it's a good word.**

I said more than I can see. Let's just look and talk.

One, AI finally wrote it right and the designer couldn't sit down

What's the worst thing that ever happened on the Ai map? Text. You tell him to write hello. It's probably Hellp, Hl10. It's like it's drunk. The GPT-image-2 directly dried the texting accuracy rate to over 99 per cent.

I made it a final exam for the fourth grade of Mathematics. The title of the scroll, the underlined under the empty title, the geometrical representation, the combination of Song italics, and the examination of the name of the school class in the sealed line are clearly and clearly arranged. At first glance, you said it was taken with a cell phone to the real exam.

Precise:

A3 edition of Math Quality Test at the end of the fourth grade of primary school in the second semester of the school year 2025, containing precise geometrical graphics, layouts of the Song body in combination with the italics title, sealed lines with the name of the school class, clear title, page numbers and scores Column

What does light print mean? I made it write the full ink, dry and wet change in ink, downmarks, paper textures, brushes, and pens.

Precise:

Generates Li Baek's clerical work in the full "Locking Roads" in the form of a ravine-style hype, rich in ink, dry and wet ink, white paper bottoms, downmarks, vertical scrolls

There's nothing wrong with commercial posters. For example, this national-tide tea poster, green rice with hot gold, paper muscle with white. All price information, door-shop activity, scrutinizing text is correct.

Precise:

3: 4 vertical versions of the new national tea and tea billboard poster, branded Yamakawa Tea, new Chinese style of luxurious restraint, coloured green, white, hot gold, image of the central cuisine cup with iced tea, cinnamon branch tilted from the top right corner, backgrounded in the inkline of the mountain, including the case of the "Mount-America" "Silent Bubble Series"

Two, it really knows what the world looks like

It's not enough to just write. What really scares the GPT-image-2 is its "knowledge of the world" - a word that listens to people: ** it really knows what things look like in the real world.**

It knows what it looks like, what the game interfaces look like, what the supermarket shelves look like. I made it a live video of the tremor, and it drew more than just a anchor, the lower left corner comment area, the right side point-sharing button, the number of viewers and running horselights, and all the UI elements were perfectly logical.

Precise:

Generates a tremor stand-up screenshot of the "EasyClaw" product, accompanied by a Labrado dog with a lobster cap with a lot of Chinese interactive text in the area of the bullet screen, with a special effect on yachts and rockets, and a background set for the live broadcast booth

Game players may feel deeper. It has also been allowed to draw images of a group of Callers in the Canyon, not only of the terrain and heroes, but also of blood strips, skill features and small map UI frames.

Precise:

Generate 16: 9 screenshots of the Renaissance Caller's Canyon game, blue perspective mid-line period, HUD full display of small maps, skill bars, equipment bars, knives, KDA scorecards, graphic engine style back

{\cHFFFFFF}{\cH00FFFF}{\cH00FFFF} Here's a tips: secretly telling you the GPT-image-2 scale is big, dddd.

Three, the design world's nuclear weapons are coming: one step at a time

Designing was a hard job. The product manager had to open Figma and the designer had to open a PS for half a day. GPT-image-2 cut the road straight off - ** from creative to finished, one sentence is enough.**

It's a particularly amazing display, it's on-line, it's on-line, it's on-line, it's off-the-shelf, it's on its own.

Phrasing (Superwide City Crossings):

A map of the city ' s long cross-sections from left to right showing morning residential areas, midday commercial areas, evening parks, night entertainment areas, a continuous street through a simple geometrical shape, with no five officials, a geometrical square, Kurzgesagt style colour, 3: 1

Phrasing (infogram of planetary guidelines for solar systems):

A vertical information map poster on the "Guide to Planets of the Solar System" with a solar system side view and a "ID" card with a diameter, rotation cycle, number of satellites next to each planet. Pluto on the edges with a small word labeling "The Ninth Planet was downgraded in 2006" and face-to-face, deep blue star sky floors, graphic design of data

Museum-level information maps are also captured:

Precise:

Generates a museum map of the cyanide plum bottle, an antique pen style, which consists of machine structure dismantling, Chinese labeling, a description of the curry material, the meaning of the tangled chords, the meaning of the sumaic cyanide, a summary of the core features, a paper sense yellow background, vertical layout

Role-setting cards, consisting of three views, face changes, equipment decomposition, palettes, world-view settings, and one step at a time in the form of animate sets.

My favorite role, for example, is the 2B role of Neil Mechanical Era, and I'm going to generate her card.

Precise:

Generate an official set-style 2B role card from the game Neil: The Mechanical Era. It contains a three-view view of the front, side, back, marked by a black Gothic dress, a black blindfold, a short white hair, and black boots. Add face change: normal look, smile look, battle anger look. The details of the "white contract" and the structure of the auxiliary device Pod 042 were broken down. Contains a complete palette: black, white, grey, gold. Add a Worldview brief box: "Forms of Sendifs No. 2-B. White background, Japanese animation set the layout style, whole image clean.

The programmer looks at an "old calendar" - the dates are right, the weeks are right, and it's time to work with AI, and I want to print it on the table.

Precise:

The design of a single page of a long-form paper, dated 21 April 2026 and early 5 March of the same year, should be "reconstructing" the "consolidated branch" and "on Friday" the "reduced library".

Four, high-energy early warning: dynasty "Friends" crossing the series

Well, it's been a long time, and it's finally on the axis of this article. This is one of the most fun series I've ever had -- making friends of the ancients.

The idea is simple: to replace all the images, state, positioning, interactions in the modern social media interface with the characters and events of the generation, but the UI framework remains unchanged.

** 5-hu16 Savannah Friends Circle:**

The "Friends of the Five-Hu" and the "SIXTEEN KINGDOMS MOMENTS" in English are ancient integration designs, social media information flow interfaces, user-images replaced with Hungla, Peng, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing, Sing.

** Qin One Circle:**

Qin Qin Qin Qin Qin Qin Qin Qin Qin Qin

** Han Silk Circle:**

"Han DYNASTY FEED" and "HUN DYAN DYNASTY FEED" crosses integration design, social media information flow interfaces, with headlines like "Han Wu, Zhang Zhui, Go to disease, Sema, " " "The River West Corridor is open" "Today's 6 o'clock "Story Updates to Wu Dy Benki" and "Hongjun" "Dung Yummen" "Hun's Court" and "Footing the West Side" with a background to draw a map of Han's bricks and silk roads, and Ui maintains a modern circle of friends.

** Don Maroon:**

"The Circle of Friends" and "TANG DYNASTY FEED" in English, "TANG DYNASTY FEED" crosses integration design, social media information flow interfaces, with headlines such as Li Bai, Yang Gui, Hyun-hyun and An Lokshan, in the state of "Hoo Ji is drunk tonight" and "The New Poe has been sent."

** Song Yachang:**

< < "Friends of the Sons" and English < SONG DYNASTY SOCIAL MEDIA FED" > > > Through the integration design, the social media flow interface was replaced by Song Dynasty human face, quill writing, which included "Today's Tea Racing" > < 3: 2 Winners in the bow party > > > > > < < < A new speech in the Vasher line > > > > The location shows "Tokyo Jae-sooty" > > "Pian Si-Huk Suty" >, "Cai-Shou-Soo-Soo-Soo-Soo-Soo-Soo-Sou-Soo-Soo-Sou-Soo-Sou-Soo-Soo-Sou-Sou-Sou-Sou-Soo-Sou-Sou-Sou-Sou-Sou-Sou-Sou-Sou-Sou-Sou-Sou-Sou-

** Won-Tun Circle:**

"The Circle of Friends" and "Yuan DYNASTY FEED" used to cross the integration design, social media information flow interfaces, with the heads Kublai, Marco Polo, Guan Hanqing, and Huang Doa, in the state of "The Grand Dum City is finished today" and "300 boats from the port of Quanzhou" and "The First Star of the Dynasty" and the location of "Ui keeps the modern circle of friends in place.

** Ming Yong Lok:**

< < < The Friends of the Great Ming and the English > > > < MING DYNASTY FEED > > > < Social media flow interface > > > with the image of Chu, Zheng and Wang Yang Ming and Tong Pak Fu > > > > Today 'Sun Dynasty > > > > < The Treasure Fleet is full > > > > > > < A shared vision of the Dragon Field > > > > > < The Namjing Qinye River ' > > > > > < Ningbo ship ship ship ship ship ship ship ship ship fragrance fragrance > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >

** 🐉 Qing Zhong Zhi:**

Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q > > > > > > > > > > > > > > > > > > > > > > The Red House Dream has been updated to 80 times today. "The West Side of the Rounder Garden has been completed." The location shows "Beijing Ant-chun Park" "Hengde Summer Hills" "Kian Ningwee House." The interaction includes "Please Ann" "Hold the Yellow Horse" "Wild Prison Report" in the background of Rong Shihing Bing and the Red Wall Yellow Wall.

V, other colored eggs (film sense/photography/classic/UI/caricature)

In addition to the friends' circle series, I've had a lot of fun, and I've done it for you:

** Film sense scene (United States subway RAW quality):**

Created an image of the American History film style, the Jewish community in the Lower East Side of New York in the 1920s, four young gang members wearing a dark suit with a soft suit, walking side by side through the smoked brick wall alley, with the long background of the steel contours of the Manhattan Bridge, a warm yellow street light flashing on the wet pebble floor, film granular sense, retro-brown tone, film-grade broad screen chart 16: 9, paying tribute to Sergio Leonenne's classic visual aesthetics.

** Rain night vibe (Tokyo lasagna):**

Tokyo Rain Night Street, a ranla shop door, neon lights on the wetland surface, cold rain lights in sharp contrast, horizons, film-class depth, 16: 9 percent, no-man vehicles focused on the focus of the door, overwritten photographs

When I saw this street view, I thought, "What else does a photographer do?"

Cultural integration contemporary design (tea cover):

Book cover design, main title "Tea" and English "The CLSIC OF TEA", ancient re-formulation of contemporary design styles, the print-printing effect of the Yen-Jing-Jing book, the texture background of the tampon, the selection of the left side of the row, the right side of the modern information map showing the six tea chromatography and the fermentation coordinates, and the bottom of the swirling print "Lande Tang Dynasty 760"

** UI Design System (Butterfly texture style):**

The visual language is extracted from the butterfly wing texture and a complete UI design system is produced that contains the home page, the mobile end card list, button components, switch controls, tab styles, very simple white base with blue and purple gradients, light shadows, Helvetica Neue font system

** Multipage Comic First Page:**

Generates the first page of 4 pages of funny comics with the title spectroscope in the first half. The title needs to be designed as an elegant type of font for interest. The image style is black-and-white line youth, reading in right order. The text in the dialogue box is left blank.

** Joint exhibition of the Nine Palace Designers:**

The nine-gauge photo matrix, each of which displays the "mechanical keyboard" of the same object designed by a contemporary well-known designer, includes the designer's head, the designer's name, the Chinese design concept, the reading of the text, the rendering of the work, the unified grid layout, the grey background.

VI. Summary: Several strengths and weaknesses that must be known

It's like GPT-image-2 in my heart:

** Advantages (really tops):**

1. Spoken to revolutionary strength. Ninety-nine per cent of the accuracy rate was not blown, and the Chinese clerical method, the UI label and the long-term layout were all off the record, and the problem of the old code had largely been resolved.
2. ** The world ' s intellectual depth is staggering**. Live interfaces, games UI, supermarket shelves, map geography, and it really knows what these things look like.
3. Step-by-step design. Information maps, UI prototypes, product packaging, posters can be delivered directly in one sentence, without the need for the Figma and PS patches.
4. ** People jump like real people**. The eye wrinkles, beards, skin senses, filaments are accurate, and the "AI plastic sense" of the old model is significantly less.
5. ** Speeds out of proportion**. The generation time of a single picture has been reduced to within 10 seconds, and the iterative experience is very slippery.

** Shortfalls (to be honest):**

1. The greyscale test is not available to everyone. Only some Plus/Pro/Team users are currently available, and free users can only stare.
2. ** The style of the illustrations is somewhat disproportionate**. According to UISDC, many illustrations are over-reinforcing in detail, somewhat like the earlier Midjourney and Niji, and the picture is a little "scatter". ** "Evidently visible" is completely invalid**. It is becoming increasingly difficult to say that the future "interceptions" are evidence of the creation of false documents, screenshots, photographs, and it is also a matter of truth.

Write at the end

In general, GPT-image-2 is not the kind of upgrading that "twices the parameters but uses them almost" and it actually fills out the most painful pits of the previous IA map - words, UI, world knowledge, realism. And it's starting to look more and more like an image engine that can go straight into real workstreams, moving from "art creation" to "business productivity tools."

The official network shows that DALL-E 2 and 3 are coming down on May 12th this year, and OpenAI has a greyscale GPT-image-2 at a point in time. They are quietly completing the old and new alternations, and the next generation of images is filled with cards.

A final reminder of friendship:** See any "interceptions" later on, with a suggestion to look more closely. In front of this model, see, it's not really true.

---

• Any of the preceding hints can be copied directly to ChatGPT (required for greyscale to GPS-image-2) to generate corresponding images. If the resulting figure is not satisfactory, it can be refined through multiple rounds of dialogue in natural languages. ♪ I'm sorry ♪