ChatGPT’s new Images 2.0 model is surprisingly good at generating text

The distinction between AI-generated images and those created by humans has become increasingly blurred, particularly with the introduction of ChatGPT's latest innovation, Images 2.0. Just a couple of years ago, AI models struggled with basic tasks, often producing bizarre concoctions like 'enchuita' or 'margartas' when asked for a Mexican restaurant menu. However, the new model now generates menus that look ready for immediate use, making it difficult for customers to spot any inaccuracies—though a $13.50 ceviche might still raise some eyebrows. Historically, AI image generators faced challenges with spelling due to their reliance on diffusion models, which reconstruct images from random noise. Asmelash Teka Hadgu, the founder and CEO of Lesan AI, explained how these models tend to focus on broader patterns rather than fine details such as text. Recently, researchers have begun exploring alternative approaches, like autoregressive models, which function similarly to language models by predicting what an image should depict. Despite inquiries during a recent press briefing, OpenAI did not disclose the exact architecture behind ChatGPT Images 2.0. However, they did highlight the model's enhanced 'thinking capabilities,' enabling it to search online, create multiple images from a single prompt, and verify its outputs. These advancements allow the model to produce diverse marketing materials and even complex multi-paneled comic strips. Moreover, Images 2.0 shows improved proficiency in rendering non-Latin scripts, including those for Japanese, Korean, Hindi, and Bengali. OpenAI asserts that this latest iteration brings a remarkable level of precision and detail to image creation. It can effectively conceptualize intricate visuals while adhering closely to user instructions, capturing fine details that often challenge previous iterations, such as small text and complex compositions—all rendered at resolutions up to 2K. While generating intricate images like comic strips may not be instantaneous, the process remains impressively quick, taking only a few minutes. Starting Tuesday, all users of ChatGPT and Codex will gain access to Images 2.0, while premium subscribers can unlock even more advanced features. In addition, OpenAI will offer access to the gpt-image-2 API, with pricing varying according to output quality and resolution.