AI Creation & Media ·
GPT image 2 is here, and the Dark Forest is here.
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
Date: 2026-04-23
Seriously, I was struck yesterday by the ability of the GPT-image-2, which was so powerful, that it was really deeply shaken. It's not just shock, it's addiction, it's input, it's input, it's direct output of real images, it's addictive.
But like technology is not wrong, it's people who are wrong, and many people are not content with creating, but rather play with ** flowers**, what's the Trumpmask live tape, A. The social statement says that Chinese users can use Claude Code, and then apologize that it's fake. Wait.
The amount of outputs produced by such maps is so large that there is a word in social media: Dark Forest.
The word does not appear for the first time. It was also mentioned when the AI painting came out. But this time, the ability of GPT-image-2 to rekindle the word and speak more seriously than ever.
The reason for this is that this photo can't really tell the truth.
What can it do
First of all, the ability, then why not this time?
GPT-image-2 is the latest image generation model for OpenAI with access to ChatGPT and API. It has several key upgrades:
**4K resolution. ** All previous AI image tools, resulting image amplification is impossible to see. GPT-image-2 supports 4K output, which can be used in print, advertising, PPT. Previously, the use of AI to generate images was jammed by resolution, and now this limit is not present.
** Text rendered. ** This is an AI image tool that has been a bad solution. An accurate translation of a text - especially in Chinese - in the image was previously a disaster-prone area. GPT-image-2 passed this off. You can add an accurate line of Chinese slogans to the image. The font, the location, the cover relationship are right.
** Multilingual support. ** Not only English, but Chinese, Japanese and Korean. OpenAI 's model training data cover multilingual text rendering, a capability that is the first full-time in the image generation field.
** Photo-level realism.** This is the most important and the most discussed. It produces human faces, skin textures, photons that are close to the camera. It's not "looks like a photo," it's "indistinguishable."
Why is this reaction stronger than ever?
AI Drawing is not new. Midjourney, Stable Diffusion, DALL-E 3, each generation is finished. But this is different.
Each of the previous generations had obvious cracks: hand paintings had collapsed, words had been coded, the light had not been natural, the skin had been too smooth as plastic. Those who are familiar with the AI image, see these details and basically judge that it was created by AI.
GPT-image-2 is different.
It creates images, none of which exist. The skin has real textures and fur holes, the light-shadow relationship is right in complex settings, and most of the time the number of fingers is right.
More importantly, it can generate ** real people in the real scene.**
It is not a fictional role in an exaggerated mask, it is not a pre-existing building, it is a photo that looks like news photography with a sense of being live, emotional and credible physical details.
What does that mean? Which means that ** anybody can be put in any scene, say anything, do anything **.
At the technical level, there are no more thresholds.
What is the Dark Forest
The concept of Dark Forest comes from Liu Zhisin's science fiction novel, The Three. The universe is a dark forest, every civilization is a hunter with a gun, and the first reaction to seeing another civilization is to destroy it - because you cannot confirm that the other is a friend.
AI's dark forest in the area of image generation means: ** When all images may be AI, trust disappears.**
You saw a picture of a public figure doing a scandal. Your first reaction was: Is this true? Now your first reaction is: could it have been AI?
The answer to this question was, "No, it's probably not, it's very expensive to fake." The answer to this question is "almost certainly a map."
One of the direct reasons for this is that,** when you send a true picture, no one will believe it, because trust costs so much that you can not believe it, rather than spend half a day identifying it.**
Bad money to drive away good money.
The problem of not keeping up with tests
It was said that the ability to generate has grown and the ability to detect will follow. This is a dynamic game, but the generation capacity is currently ahead of it.
Technical tests have several lines of thought: looking at metadata, analysing statistical characteristics at pixel level, and trace model fingerprints. But each has its limitations. Metadata can be deleted or falsified, pixel-level features become less visible after they are produced and trace fingerprints are difficult to maintain after models have been iterative.
Even more problematic is the fact that, in practice, images are transcribed, compressed, relayed, and edited twice. Every step, the remaining AI features are a little less, and by the time the images reach you, they may not be detectable.
So the reality is that the generation end is running, the test end is chasing, and it's pushing hard.**
The real situation of the creator
Finish the macro and say micro.
The emergence of GPT-image-2 has changed the whole ecology for the creators who are actually using the AI image tool.
The boundaries of these professions have been significantly reduced by illustrations, designers and photographers. The drawings, which took three days, are now produced in GMT-image-2 for five minutes. Although quality may not be fully professional, ** is good enough ** This threshold has been crossed.
Good enough means that a lot of business scenes that don't require a top level can already be met by AI.
Advertising companies, media, content platforms - where the quality of images is required to be "accurate, good-looking, usable" rather than "artistic". After the GPT-image-2 has reached this standard, a large number of intermediate requirements will no longer need to be met by real people.
And just yesterday, I was told that he had read my article and found that GPT-image-2 was powerful and that it would no longer be necessary to pay for design.
This is not to say that the top creators will be replaced, but to say that ** the middle layer of the employment structure is quickly removed**.
Technology itself has no position. GPT-image-2 can be used as a poster, an educational content, a prototype of a product, as well as to forge evidence, impersonate identity and create false information.
This is not a matter for technology to decide, but for society to face.
However, there is one track-off that cannot be avoided:** Limiting capacity is tantamount to limiting possibilities.**
GPT-image-2 capabilities are double-edged swords. For creators is the liberation of productivity; for malicious actors is the magnification of destructive power. It is almost impossible to try to keep only the good half.
Water's pouring out. What we can do is figure out how the water will flow and figure out where it can affect.
GPT-image-2 is a node in the AI image generation field.
The question before is "Is AI capable of generating images that look real?" The problem is now largely resolved.
The next question is: ** How can trust be created when all the images may be AI?**
It's not a technical issue. It's a social issue. New certification mechanisms, new trust infrastructure, new self-media literacy are needed.
Dark forests are not a state that can be eliminated. It is a reality that must be faced after AI has developed imagery to this stage.
What can be done is not to rekindle the forest, but to find new ways of walking in the dark.
Also, give you a few Prompt and Vincent charts.
Photography style
Photo of an extremely micro-photographic photograph of a bee parked on a purple lavender, visible on the eye and feathers on the wings, with a vanity of the background into a dreamy green spot, 2: 1 magnification rate
to create a deep-space astronomical photo image of Orion's big nebula M42, woven with red and blue nebula from ionizing hydrogen, centered in bright four stars, full of stars, the Hubble Space Telescope style coloring Board
Takes a long exposure photograph, at night on the coast, waves smug the rocks into silky white fog, clouds in the sky with dynamic drag effects, ND1000 filtering sense.
Artism
{\cHFFFFFF}{\cH00FFFF}{\cH00FF00} Creates a stereoist Stilling painting, a violin, a glass and a musical spectroscopy, with multiple perspectives.
{\cHFFFFFF}{\cH00FF00} Creates a superrealistic digital art that floats through an open wooden door in an empty desert with a deep sky and a huge eye, a melting clock on the ground and a combination of El Salvador and René Magrat.
{\cHFFFFFF}{\cH00FFFF}{\cH00FFFF} Create a very concise landscape, a horizontal line that divides the picture into a grey blue sky and a dark gray ocean, with only one canoe in the lower left corner, with a large amount of white left, Donald Judd and Agnes Martin style.
Create a digital portrait of a broken art style, half-faced characters with RGB channel separation and pixel torn apart, background with horizontal scanning line snowflakes, data in Sabor format damaging aesthetics
Create a low polygon-style island scene, sunset time, geometrical mountain ranges, pine trees and water reflections, lightened colours, visible polygonal grids, like memorial valley games
Design a very brief musical poster. The event was called "Audience Valley Music Festival". The main vision is a black pebble sank to the surface, with large amounts of white above it, with only thin Song and English in the layout.
Sebastian
{\cHFFFFFF}{\cH00FFFF} Creates a steam punk-style design of a Victorian complex of brass gears, nailed steel plates, leather balloons and wooden propellers floating over London during the Industrial Revolution.
Producing a video cover of a 1990s VCD rental store called "Death Code ", starring Zhou Yunfa, Lin Qingxia. The cover is crowded, full of action-film explosive elements, with the words "maisal laser" on the bottom, and a slight scanned line wear. Sensor.
{\cHFFFFFF}{\cH00FF00} Create a 16-bit pixel art-style video game with a sword-bearer standing on the edge of the cliff, far away with pixel castles and floating clouds, colourful, retropy games Sensor.
