The Image Generation Model That's Hotter Than Nano Banana Has Leaked, Screenshots Are No Longer Evidence | Includes Prompts

marsbitPublicado em 2026-04-19Última atualização em 2026-04-19

Resumo

A new AI image generation model, widely referred to as "GPT Image 2," has been leaked and is demonstrating significant advancements over predecessors like DALL-E 3 and even Google's Nano Banana Pro. It excels in four key areas: text rendering, prompt adherence, photorealism, and world knowledge. The model can generate highly accurate text in multiple languages, including complex Chinese characters, making it capable of producing convincing fake documents, UI screenshots, and product labels. This capability also raises concerns about the reliability of using screenshots as evidence. The model is currently in A/B testing, with a full release expected around May 2026 when DALL-E services are officially retired. It is accessible for testing on the LM Arena platform. The article includes several prompt templates optimized for the model, such as generating realistic app screenshots, product photos with detailed labels, and street scenes with accurate signage. This advancement is reshaping creative workflows but also accelerating the displacement of some traditional design roles.

Is your impression of text-to-image still stuck on Nano Banana?

But kid, times have changed again.

@johnAGI168 https://x.com/johnAGI168/status/2044781168151724067

@0115hippo https://x.com/0115hippo/status/2044722124611539160

In early April, three anonymous image models, codenamed maskingtape-alpha, packingtape-alpha, and gaffertape-alpha, appeared on the LM Arena evaluation platform. They disappeared a few hours later.

OpenAI has not officially announced this model yet, but based on the metadata returned by the API and user-side testing records, it has already gained a widely accepted name: GPT Image 2.

Screenshots Can No Longer Be Used as Evidence

Over the past few years, one of the most obvious weaknesses of AI image generation models has been text within images. In the DALL-E 3 era, if you asked it to write "Hello" in an image, it might output "Hellp" or even "Hl10", with letters tilting drunkenly. GPT Image 1 improved a lot, handling simple English labels. By GPT Image 1.5, its accuracy in rendering English text was close to 95%, but it still had significant flaws with non-Latin scripts like Chinese, Japanese, and Korean.

But the leaked sample images from GPT Image 2 have changed this impression.

@MrLarus https://x.com/MrLarus/status/2044824800909054181

@akokoi1 https://x.com/akokoi1/status/2044789531615056175

The text in the images is exactly what it should be. Chinese characters are clear, with accurate glyphs and complete strokes. Someone tested generating an ID card-style image, where the name, address, and ID number were all rendered correctly, with neat formatting, looking at first glance like a photo of a real document.

This is good news. The improvement in text rendering means generating infographics, posters, product packaging, and complex charts becomes more reliable.

But there's always another side to the coin. A model that can generate photo-realistic ID-style images and precisely render UI screenshots naturally makes "screenshots can be used as evidence" increasingly questionable.

By comparison, this is also a core difference between the GPT Image series and other models. Midjourney still has no progress in text rendering, and the Stable Diffusion series also has this old problem. According to the leaked Arena test results, GPT Image 2 surpassed Midjourney in four dimensions: text rendering, instruction following, photorealism, and world knowledge. Midjourney's advantages are mainly retained in artistic style and aesthetic control.

Does It Really Know What the World Looks Like?

A tester asked the model to generate a hypothetical GPT-8 product pricing page. The resulting image had a layout that was indeed in the style of the OpenAI website, with button placement and font choices resembling those from a real interface, and the hierarchical logic of the price table was correct.

GPT Image 2 can generate images extremely similar to real software interfaces, including browser windows, mobile app interfaces, and data visualization charts, with a level of fidelity unmatched by the previous generation.

@johnAGI168 https://x.com/johnAGI168/status/2044781168151724067

@levelsio https://x.com/levelsio/status/2040333489476681758

This will lead to some very interesting practical uses. When designers are creating product prototypes, they don't need to open Figma first and draw a bunch of wireframes; they can directly describe the desired interface in text, and the output is a reference image that can be used for team discussions. When creating investor decks, they can show a "product screenshot" without waiting for an engineer to write code. When writing documentation, example interface images for illustration can be generated directly, without having to think about where to find screenshots for a blank page.

@marmaduke091 https://x.com/marmaduke091/status/2040338311873515597

Image Generation Is No Longer Just "Image Generation"

OpenAI has already announced that DALL-E 2 and DALL-E 3 will officially cease service on May 12, 2026. Azure OpenAI's DALL-E 3 was retired early in February.

DALL-E was the first place many people encountered AI image generation, from those blurry early works to today, in just a few short years.

Meanwhile, Google, which had just established its industry position with Nano Banana Pro in early 2026, might feel the pressure. Early test reports indicate that GPT Image 2 simultaneously surpasses Nano Banana Pro in three dimensions: realism, text rendering, and world knowledge. This kind of triple win is not common.

For creators, the feeling is complex. Illustrators, graphic designers, and photographers are not facing this topic for the first time. Since the release of GPT Image 1, the number of freelance graphic design positions has decreased by about 18%. AI has indeed replaced the decision to "hire someone to do this" in certain scenarios, but it is also creating new ways of working, allowing one person to do more.

The evolution speed of image generation models no longer leaves much time for adaptation. It was only a few months from GPT Image 1's launch to version 1.5. And from 1.5 to 2, it's only been about half a year. Each generation solves the core shortcomings of the previous one while opening up new possibilities.

GPT Image 2 is currently still in the A/B testing phase, with some ChatGPT users randomly gaining access. The official release window is widely predicted to be around May, coinciding with the retirement of DALL-E. If you want to experience it early, you can currently try your luck on the LM Arena evaluation platform.

Test Address: https://arena.ai

Based on community feedback and the known strengths of this model, the following prompt templates can maximize your chances of success:

UI/Screenshot Prompt: A photorealistic screenshot of a mobile banking app, clearly showing transaction history with dates, amounts, and merchant names legible. iPhone 16 screen, natural hand holding the phone, coffee shop background.

Product Label Prompt: A photographic product photo of a craft beer bottle, with clear label details showing the brewery name "Oakridge Brewing Co.", alcohol content 6.8%, a mountain logo, and an ingredient list. Studio lighting, white background.

Signage Prompt: A street scene photo of a Tokyo alley at night, showing multiple neon signs in both Japanese and English, including a ramen shop sign reading "Ichiban Ramen — Est. 1987", a karaoke bar sign, and various glowing advertisements. Wet, reflective pavement with light reflections.

Interface/World Knowledge Prompt: A photorealistic YouTube video screenshot showing a video titled "How to Assemble a Computer in 2026" with 2.3 million views, featuring realistic comments, sidebar video recommendations, and channel info. Desktop browser view.

Widescreen Trigger Prompt: A cinematic widescreen photo of an IKEA store exterior at dusk, showing the glowing IKEA sign, a parking lot with realistic cars, and shoppers entering and leaving. Golden hour lighting, 16:9 format.

Unattributed image sources and references: https://miraflow.ai/blog/how-to-use-duct-tape-ai-model-arena-gpt-image-2-guide

This article is from the WeChat public account "APPSO", author: Discovering Tomorrow's Products

Perguntas relacionadas

QWhat is the name of the leaked image generation model mentioned in the article, and what is its significance?

AThe leaked model is referred to as GPT Image 2. Its significance lies in its dramatic improvement in text rendering accuracy, especially for non-Latin scripts like Chinese, and its ability to generate highly realistic images, including convincing UI screenshots and document-style images, which challenges the reliability of screenshots as evidence.

QHow does GPT Image 2's performance compare to other models like Midjourney and Google's Nano Banana Pro?

AAccording to the article, GPT Image 2 outperforms Midjourney in text rendering, prompt following, photorealism, and world knowledge, with Midjourney retaining an advantage mainly in artistic style and aesthetic control. It also reportedly surpasses Google's Nano Banana Pro in realism, text rendering, and world knowledge.

QWhat are some of the potential practical applications of GPT Image 2's capabilities?

APotential applications include generating product prototypes and UI mockups for designers, creating realistic 'screenshots' for investor decks without coding, producing example interface images for documentation, and generating accurate product labels, packaging, and information graphics.

QWhat major change is OpenAI making to its image generation services in relation to this new model?

AOpenAI has announced that DALL-E 2 and DALL-E 3 will officially stop service on May 12, 2026, with Azure's DALL-E 3 having already been retired in February. This suggests a transition to the new GPT Image model series.

QWhere can users currently try to access or test the GPT Image 2 model, and what is a recommended strategy for getting good results?

AThe model is currently in A/B testing, with some ChatGPT users randomly gaining access. Users can also try their luck on the LM Arena评测平台 (arena.ai). The article recommends using specific, detailed prompt templates focused on UI/screenshots, product labels, signage, interface/world knowledge, and widescreen formats to maximize success.

Leituras Relacionadas

Apple Also Has to Pay Rent Now

Apple Pays Rent Too: The Two-Way Flow of "Traffic Tax" and "AI Capability Rent" Between Tech Giants For over two decades, Google has paid Apple an estimated $20 billion annually to remain the default search engine on Safari, a "traffic tax" for a critical user entry point. However, in 2026, the direction of this cash flow partially reversed. Apple agreed to pay Google roughly $1 billion per year to license its Gemini AI models, as Apple's own models reportedly struggled with complex tasks. This creates a unique dynamic: Apple acts as the "landlord" in the established search ecosystem, collecting rent from Google for access. Simultaneously, in the emerging AI arena, Apple becomes the "tenant," paying Google for access to cutting-edge AI capabilities it cannot currently match internally. While Apple claims its new models are "distilled" from Gemini outputs and contain "not a drop" of Google's original code, core dependencies remain. Its knowledge base is refined using Gemini's outputs, and its most powerful cloud model runs on Google's infrastructure. Apple has structured the deal as non-exclusive, allowing it to theoretically switch AI suppliers—a hedge against over-reliance. The future hinges on whether advanced AI models become a commodity (cheap and abundant) or remain a concentrated, scarce resource (expensive and controlled by few). Apple is betting on the former, leveraging its massive device ecosystem to be a powerful, choosy customer. If the latter proves true, its bargaining power could erode. This power dynamic is extending to developers. Apple, Google, and WeChat are all pushing for apps to expose their core functions as standardized "actions" or "intents" that their respective AI assistants (Siri, Gemini, WeChat AI) can directly call. The new scarce resource is no longer just app store visibility, but "being selected by the AI." The currency of "rent" has changed from a 30% revenue share to ceding control over how users interact with an app's functions.

marsbitHá 8m

marsbitHá 8m

Missed the SpaceX IPO? WEEX's "First Trade Protection" Lets You Experience US Stock Trading Risk-Free.

With the excitement around SpaceX's recent public listing reigniting interest in the US stock market, Chinese investors face significant challenges accessing compliant and convenient trading channels following regulatory actions against major online brokers. This article explores the available options, highlighting their risks and limitations. Traditional paths for US stock investments remain problematic. Qualified Domestic Institutional Investor (QDII) and Listed Open-Ended Fund (LOF) products, while compliant, suffer from high fees, significant purchase premiums, and a very limited selection of assets. Small, unregulated offshore brokers pose substantial risks, including potential insolvency. While secure, VIP accounts at banks in Hong Kong or Singapore require high minimum deposits (often 1-2 million RMB) and in-person visits, placing them out of reach for most retail investors. The article positions cryptocurrency exchanges, specifically their TradFi (traditional finance on-chain) offerings, as a compelling alternative. Platforms like WEEX are noted for providing access to a wide range of US stocks and ETFs, including SpaceX (SPCXON), through tokenized assets. This method offers advantages such as a single account for both crypto and traditional assets, USDT-based settlement avoiding fiat complexities, flexible leverage, and robust risk management. To attract users, WEEX is promoting a "First Trade Guarantee" campaign. Running from June 15 to July 8 (UTC+8), it features a $30,000 prize pool. Users who trade $500 worth of US stock contracts can qualify for a guarantee on their first eligible trade: 100% loss coverage up to $30 or a 20% bonus on profits up to $30. The campaign is presented as a low-risk opportunity for both crypto natives and traditional investors to experience US stock trading.

marsbitHá 10m

Missed the SpaceX IPO? WEEX's "First Trade Protection" Lets You Experience US Stock Trading Risk-Free.

marsbitHá 10m

How Difficult is Chip Making? A Division Error Costs 475 Million Dollars

How Hard Is It to Make a Chip? A Division Error Cost $475 Million Chip expert Shi Kan, a researcher at the Chinese Academy of Sciences and a popular tech creator, explains the immense challenges of chip development. Chips are foundational to modern technology, but their creation is extraordinarily difficult. The journey from sand to a functional chip involves complex design and manufacturing, but a critical bottleneck is verification—ensuring the design works flawlessly before costly production. A single, undetected bug can have catastrophic consequences, as illustrated by the infamous 1994 Intel Pentium FDIV bug. A flaw in the floating-point division unit forced a recall costing $475 million. Unlike software, chips cannot be easily patched after manufacture, making "first-time success" paramount. However, industry surveys show only 24% of chip projects achieve this; over three-quarters require at least one costly re-spin due to design flaws. Verification has thus become the dominant phase, consuming up to 70% of the design cycle. The core challenge is a "verification impossible triangle" between high performance, good debuggability, and low cost. Exhaustively verifying a modern CPU core could take 15,000 years with software simulation, or 30 years with advanced hardware emulation—timeframes utterly impractical for development. Despite being essential, verification is often seen as unglamorous "dirty work," receiving less academic attention than fields like AI. Shi and his team are tackling this by developing an agile verification research framework called ENCORE, based on FPGA technology, to improve verification efficiency and debug capability. Beyond research, Shi engages in public science communication through long-form video content, aiming to demystify chip technology, AI, and computer science. He argues for the value of pursuing "hard and long-term" endeavors, whether in the meticulous world of chip verification or in creating substantive educational content, believing such sustained effort is likely the right path forward.

marsbitHá 19m

How Difficult is Chip Making? A Division Error Costs 475 Million Dollars

marsbitHá 19m

Claude to Mandate "Face-Scan ID Verification", No ID No Service Starting July?

Anthropic, the creator of Claude AI, has sent privacy policy update emails to users, signaling a significant shift. The key change, effective July 8, is the potential requirement for consumer-level users (Free, Pro, Max plans) to verify their age or identity. This verification would be conducted through the third-party service Persona, involving uploading a government-issued photo ID and taking a live selfie for comparison. Anthropic states this data is for security purposes only, will not be used for model training, and is processed by Persona, not stored on its own servers. The update also clarifies data handling for Claude's new capabilities: when performing multi-step tasks or connecting to third-party apps, user data may flow between Anthropic and those external services. Additionally, more information may be collected from users who participate in Anthropic research. This move is seen as a major step towards establishing accountability as AI agents become more powerful and autonomous, capable of executing complex, real-world tasks. It follows previous enforcement actions, like the banning of the "Fable 5" account, and indicates a broader industry trend toward stricter user identification and safety measures. The verification is expected to apply in specific scenarios, particularly as users engage Claude in more complex agentic workflows.

链捕手Há 32m

Claude to Mandate "Face-Scan ID Verification", No ID No Service Starting July?

链捕手Há 32m

Blockchain Has Finally Started to Sail into the Mainstream After 18 Years

Blockchain Finds Its True Path After 18 Years: Becoming the Financial Backbone for AI Agents and Autonomy This analysis explores a pivotal shift in the blockchain and crypto investment landscape, driven by the dominance of AI. Major venture capital firms, including Variant, Paradigm, Haun Ventures, and YZi Labs, are moving beyond pure "crypto" investment theses. They are expanding their focus to AI, robotics, and frontier tech, signaling that blockchain is no longer seen as a standalone sector but as an underlying infrastructure layer. The core argument is that blockchain's killer application may not be user-facing apps, but rather providing the economic rails for the coming wave of AI agents, autonomous robots, and automated systems. Key capabilities like self-custody wallets, programmable stablecoins for micropayments, on-chain identity, and verifiable smart contracts are positioned as essential for a future where machines conduct economic activity. The recent $1.4 billion investment by Tether (via its venture arm) in German robotics company NEURA Robotics exemplifies this, aiming to embed Tether's wallet tools directly into robots for autonomous transactions. While many "AI + Crypto" projects remain superficial, the article concludes that true value lies where crypto is a necessary component—enabling machine-to-machine payments, agent autonomy, verifiable data provenance, and open financial settlement for the AI era. For crypto venture capital, this convergence with AI represents both an adaptation to shifting capital flows and a potential path to unlocking the large-scale, non-speculative utility the industry has long sought.

marsbitHá 40m

Blockchain Has Finally Started to Sail into the Mainstream After 18 Years

marsbitHá 40m

Trading

Spot

Futuros

Artigos em Destaque

Como comprar BANANA

Bem-vindo à HTX.com!Tornámos a compra de Banana Gun (BANANA) simples e conveniente.Segue o nosso guia passo a passo para iniciar a tua jornada no mundo das criptos.Passo 1: cria a tua conta HTXUtiliza o teu e-mail ou número de telefone para te inscreveres numa conta gratuita na HTX.Desfruta de um processo de inscrição sem complicações e desbloqueia todas as funcionalidades.Obter a minha contaPasso 2: vai para Comprar Cripto e escolhe o teu método de pagamentoCartão de crédito/débito: usa o teu visa ou mastercard para comprar Banana Gun (BANANA) instantaneamente.Saldo: usa os fundos da tua conta HTX para transacionar sem problemas.Terceiros: adicionamos métodos de pagamento populares, como Google Pay e Apple Pay, para aumentar a conveniência.P2P: transaciona diretamente com outros utilizadores na HTX.Mercado de balcão (OTC): oferecemos serviços personalizados e taxas de câmbio competitivas para os traders.Passo 3: armazena teu Banana Gun (BANANA)Depois de comprar o teu Banana Gun (BANANA), armazena-o na tua conta HTX.Alternativamente, podes enviá-lo para outro lugar através de transferência blockchain ou usá-lo para transacionar outras criptomoedas.Passo 4: transaciona Banana Gun (BANANA)Transaciona facilmente Banana Gun (BANANA) no mercado à vista da HTX.Acede simplesmente à tua conta, seleciona o teu par de trading, executa as tuas transações e monitoriza em tempo real.Oferecemos uma experiência de fácil utilização tanto para principiantes como para traders experientes.

141 Visualizações TotaisPublicado em {updateTime}Atualizado em 2026.06.02

Discussões

Bem-vindo à Comunidade HTX. Aqui, pode manter-se informado sobre os mais recentes desenvolvimentos da plataforma e obter acesso a análises profissionais de mercado. As opiniões dos utilizadores sobre o preço de BANANA (BANANA) são apresentadas abaixo.