OpenAI Upgrades Image Generator for Developers - DALL-E 3

OpenAI Extends Advanced Image Generation to Developers via API

On Wednesday, OpenAI made the technology powering its enhanced image generation feature within ChatGPT accessible to developers through its API. This allows for the integration of this capability into a wide range of applications and services.

Viral Launch and Increased Demand

The new image generator, released to most ChatGPT users in late March, quickly gained popularity. Its ability to produce remarkably realistic images, including those in the distinctive Ghibli style and "AI action figures," led to widespread attention.

This surge in interest presented both opportunities and challenges for OpenAI. While it resulted in millions of new ChatGPT sign-ups, it also significantly increased the strain on the company’s infrastructure. Over 700 million images were created by more than 130 million ChatGPT users within the first week of the tool’s release, as reported by OpenAI.

Introducing gpt-image-1

The image-generation functionality within the OpenAI API is driven by an AI model known as “gpt-image-1.” This model is inherently multimodal, meaning it can generate images in diverse styles.

gpt-image-1 is capable of adhering to specific instructions, utilizing existing knowledge, and accurately rendering text within images.

Developer Control and Flexibility

Developers utilizing gpt-image-1 can generate multiple images simultaneously. They also have control over the quality of the generated images, which directly impacts the speed of creation.

Safety and Moderation

OpenAI confirms that gpt-image-1 incorporates the same safety measures as image generation within ChatGPT. These include restrictions preventing the creation of content that violates the company’s policies.

Developers can adjust the moderation sensitivity, choosing between “auto” for standard filtering or “low” for reduced restrictions. According to OpenAI documentation shared with TechCrunch, the “low” setting allows for fewer limitations on potentially unsuitable content.

Authenticity and Watermarking

All images generated using gpt-image-1 are embedded with C2PA metadata watermarks. This allows supported platforms and applications to identify them as AI-generated.

Pricing Structure

The pricing for gpt-image-1 is as follows: $5 per million input tokens for text and $10 per million input tokens for images, with $40 per million output tokens for images. (Tokens represent the fundamental units of data processed by the model.)

This translates to approximately 2 cents, 7 cents, and 19 cents per generated image for low-, medium-, and high-quality square images, respectively, according to OpenAI’s calculations.

Early Adopters

Several companies are already integrating or testing gpt-image-1, including Adobe, Airtable, Wix, Instacart, GoDaddy, Canva, and Figma.

For example, Figma’s Figma Design platform now enables users to generate and edit images using gpt-image-1. Instacart is currently evaluating the model for use in creating images for recipes and shopping lists.