OpenAI has implemented robust safeguards to address safety and misinformation concerns, embedding C2PA metadata in generated images to identify them as AI-created. However, this metadata can be easily removed, especially on social media platforms. Currently, the image generation feature is exclusive to OpenAI's $200 per month Pro subscription, with plans to expand access to Plus and free users soon.
Key takeaways:
- OpenAI has introduced new image generation capabilities for ChatGPT, using the GPT-4o model, which improves text rendering and follows instructions better.
- The GPT-4o model uses an autoregressive approach for image generation, unlike the diffusion model used by DALL-E, and is fine-tuned for more photorealistic images.
- Despite improvements, the model still struggles with small lettering, non-Latin scripts, and may hallucinate information, raising safety and misinformation concerns.
- Currently, GPT-4o image generation is available only to subscribers of OpenAI's $200 per month Pro subscription, with plans to expand access to more users in the future.