DALL·E 3 Integration in ChatGPT Plus and Enterprise
OpenAI has released DALL·E 3, now available for ChatGPT Plus and Enterprise users. This model represents a significant advancement in image generation, offering improved visual quality, higher detail, and a stronger ability to follow complex, detailed prompts compared to its predecessor.
Technical Advancements in Prompt Adherence
DALL·E 3 achieves superior prompt adherence by utilizing a state-of-the-art image captioner to generate high-quality textual descriptions for the training data. By training on these improved captions, the model has become significantly more attentive to user-supplied captions, allowing it to reliably render intricate details such as text, hands, and faces.
Visual Capabilities and Formatting
DALL·E 3 produces images that are more visually striking and crisper in detail than previous versions. The model supports multiple aspect ratios, including both landscape and portrait orientations, enabling users to create images with flexible formatting requirements.
Safety Systems and Content Moderation
OpenAI employs a multi-tiered safety system to prevent the generation of harmful imagery, including adult, violent, or hateful content. This system operates by running safety checks on both the initial user prompts and the resulting images before they are delivered to the user.
To further refine these systems, OpenAI collaborated with expert red-teamers and early users to identify and only address gaps in coverage. This process specifically targeted edge cases for graphic content, such as sexual imagery, and tested the model's ability to generate misleading images.
Artist and Public Figure Protections
As part of the deployment preparation, OpenAI implemented measures to limit the model's likelihood of generating content in the style of living artists or images of public figures. The company also focused on improving demographic representation across the generated images.
AI Provenance and Detection
OpenAI is evaluating an internal provenance classifier designed to identify whether an image was generated by DALL·E 3. Internal evaluations show the following accuracy rates:
- Unmodified images: Over 99% accuracy in identifying DALL·E generated images.
- Modified images: Over 95% accuracy when images undergo common modifications such as JPEG compression, resizing, cropping, or the superimposition of real image cutouts.
While these results are strong, OpenAI notes that the classifier currently provides a likelihood of generation rather than a definitive conclusion. This tool is part of a broader effort to collaborate across the AI value chain to help users distinguish AI-generated visual content from real imagery.