OpenAI Revolutionizes Image Creation with ChatGPT Images 2.0

OpenAI Enriches ChatGPT Images 2.0 with New Features

OpenAI recently unveiled a major update to its image generator, called ChatGPT Images 2.0. This new model incorporates advanced reasoning and web search capabilities, allowing for the creation of more coherent and varied images. Users can now generate up to eight images from a single prompt, with significantly improved text handling, including those using non-Latin scripts.

An Official and Innovative Image Model

OpenAI's image model, now official, is based on the new GPT Image 2. This model shares similarities with Google's Nano Banana Pro, particularly its ability to "think" before producing images. This reflection process, which may include web searches, aims to enhance the accuracy and diversity of generated images. However, these advanced features are reserved for subscribers of the ChatGPT Plus, Pro, and Business plans.

Multiple and Coherent Image Generation

Thanks to the reflection mode, ChatGPT Images 2.0 can produce up to eight coherent images from a single prompt. OpenAI provides usage examples such as creating one-page mangas, graphic series for social media, or design plans for different rooms in a house. These capabilities help maintain consistency in characters, objects, and styles across the various generated scenes.

Improved Image Quality for All

Regardless of the use of reflection mode, all ChatGPT users benefit from enhanced image quality. OpenAI claims that the new generator captures photo characteristics more accurately and improves the quality of pixel art, mangas, movie images, and other types of visuals. The model is also optimized to handle complex elements such as small text, iconography, and dense compositions.

Extended Support for Formats and Resolutions

The model supports a variety of aspect ratios, ranging from 3:1 (ultra-wide) to 1:3 (ultra-tall), allowing for the creation of formats suitable for banners, presentation slides, and mobile screens. Image resolution can reach up to 2K via the API, providing increased possibilities for developers and content creators.

Flexible Pricing Based on Tokens

Developers can integrate the model into their products via the API under the name gpt-image-2. OpenAI offers token-based pricing, with costs of $8 per million input image tokens and $30 per million output image tokens. Prices vary based on the quality and resolution of the images, ranging from $0.006 for a low-quality image to $0.211 for a high-quality image at 1024 x 1024 pixels.

Larger Resolutions at Lower Costs

At higher resolutions, GPT Image 2 proves to be more economical than its predecessors. For example, a high-quality image of 1024 x 1536 pixels costs $0.165, compared to $0.20 for GPT Image 1.5. However, at the standard resolution of 1024 x 1024 in high quality, the new model is slightly more expensive at $0.211.

Diverse and Promising Use Cases

OpenAI highlights several use cases for its new model, including localized advertising, infographics, educational content, and design tools. In Codex, image generation will be integrated directly into the workspace, simplifying access for users.

Impressive Performance in Tests

In internal tests, ChatGPT Image 2 demonstrated high efficiency. Both instant and reflection modes successfully managed complex prompts with particular attention to detail. For instance, a hyper-realistic image of a monkey holding a pink banana on a tiger, with a horse riding an astronaut, was rendered with remarkable clarity and precision.

Limited Access and First Impressions

OpenAI's image model, gpt-image-2, is currently being tested by select ChatGPT users. The first generated images, which closely resemble real photographs, have been shared on X and Reddit. For now, access appears to be restricted to testers based in the United States.

A Leap Forward in Complex Image Generation

GPT-Image 2 excels at creating complex images and diagrams with text, making it an ideal choice for advertising and educational applications. OpenAI plans to officially unveil this model during a livestream, promising to address typical imperfections of previous versions, such as overly smooth skin and unrealistic lighting.