ChatGPT’s modern image generator can now produce readable text inside images, a capability that previously eluded AI models and often resulted in nonsensical labels like “enchuita” on a taco menu.
This advancement marks a significant shift in how multimodal AI systems handle fine-grained details, particularly in rendering legible characters and symbols that were historically lost in the noise-reconstruction process of diffusion models.
According to testing by TechCrunch, the updated model — referred to internally as Images 2.0 — successfully generated a Mexican restaurant menu that could pass as authentic, with correct spelling and realistic pricing, aside from one conspicuously high-priced ceviche at $13.50 that raised eyebrows about ingredient quality.
Two years prior, attempting the same task with DALL-E 3 yielded garbled inventories of fake dishes, underscoring how far image generation has come in mastering textual fidelity within visual outputs.
The improvement stems from a shift away from pure diffusion techniques toward architectures that behave more like language models, enabling better prediction and reconstruction of spatial text elements.
Asmelash Teka Hadgu, founder and CEO of Lesan AI, explained to TechCrunch in 2024 that diffusion models struggle with text because written elements occupy few pixels, causing the model to prioritize broader visual patterns over minute linguistic details.
Researchers have since explored autoregressive alternatives, which model image generation sequentially — much like how LLMs predict the next word — allowing for greater precision in structured outputs such as labels, icons, and user interface components.
Although OpenAI did not disclose the exact architecture behind Images 2.0 during a recent press briefing, the company confirmed the integration of “thinking capabilities” that allow the model to search the web, reason about image structure before rendering, and generate multiple variations from a single prompt.
These capabilities enable users to produce complex, multi-panel outputs like comic strips or marketing assets in various dimensions while maintaining consistency in characters, style, and layout across frames.
The Verge reported that when thinking mode is activated, the system can pull real-time information from the web, analyze uploaded files, and generate up to eight coherent images simultaneously — a feature aimed at designers, educators, and content creators needing scalable visual storytelling.
OpenAI emphasized that Images 2.0 now supports resolutions up to 2K and a broader range of aspect ratios, from wide 3:1 banners to tall 1:3 mobile-first layouts, expanding its utility beyond square-format social media posts.
Crucially, the model shows marked improvement in rendering non-Latin scripts, including Japanese, Korean, Chinese, Hindi, and Bengali — languages where precise character formation is essential for readability and cultural accuracy.
This multilingual strength addresses a longstanding gap in AI image generation, where non-Western scripts were often distorted or illegible, limiting global usability.
The model’s training data cutoff is December 2025, meaning it may lack awareness of events, trends, or terminology emerging afterward — a constraint that could affect prompts involving recent product launches, cultural moments, or breaking news.
Despite this, OpenAI positions Images 2.0 as a tool for professional workflows, asserting that it can follow detailed instructions, preserve user-specified elements, and render fine details such as small text, icons, and UI components that previously caused failures in other models.
While generation is not instantaneous — complex outputs like multi-paneled comics still capture several minutes — the latency is deemed acceptable for professional apply where precision outweighs speed.
Access to Images 2.0 is being rolled out to all ChatGPT and Codex users, with thinking features available to subscribers on Plus, Pro, Business, and Enterprise tiers.
The rollout reflects OpenAI’s broader strategy of integrating multimodal reasoning into its ecosystem, blurring the line between text and image generation through shared underlying capabilities.
As AI-generated visuals become increasingly indistinguishable from human-created content in professional contexts, the implications for design, advertising, and education are poised to grow — though questions remain about attribution, consent, and the potential for misuse in generating deceptive or misleading visuals.
How does Images 2.0 improve text rendering in images compared to earlier models?
Images 2.0 moves beyond diffusion models by using architectures that predict image elements more like language models, allowing it to render small text, icons, and UI elements with greater accuracy — a task earlier models struggled with due to their noise-reconstruction approach.
What are the limitations of Images 2.0 despite its improved capabilities?
The model’s knowledge is current only up to December 2025, which may reduce accuracy for prompts involving recent events, and while it generates detailed images, the process is not instantaneous — complex outputs like multi-panel comics still take several minutes to complete.
