Introducing 4o Image Generation

Source: OpenAI Date: 2025-03-25 URL: https://openai.com/index/introducing-4o-image-generation

Summary

OpenAI’s March 2025 launch of a new image generation capability built directly into GPT-4o — distinct from DALL-E 3, this is native image generation from the multimodal GPT-4o model rather than a separate image model. The result: more coherent image generation that understands context from the full conversation, can incorporate text accurately into images, and handles complex multi-element compositions better than DALL-E 3. The launch went viral quickly due to Studio Ghibli-style image generation that swept social media.

Implications

Native multimodal image generation. The architectural shift from separate image model (DALL-E) to unified multimodal model (GPT-4o generating images natively) is significant. The model understands the conversation context when generating images — it can reference earlier text, maintain character consistency across generations, and incorporate nuanced requests more accurately. This is a qualitative capability improvement, not just quality tuning.

The Ghibli moment. The viral Studio Ghibli-style image generation was simultaneously a product success (massive adoption spike), a policy conversation (IP implications of style emulation at scale), and a competitive pressure moment (Midjourney, Adobe, Stable Diffusion all faced immediate “can you do Ghibli-style?” comparisons). OpenAI’s handling of the IP question — generating in the style of a studio without explicit licensing — set a precedent other labs had to navigate.

Thread: image generation evolution. The successor to DALL-E 3 (October 2023) as the primary image generation in ChatGPT. The new ChatGPT Images product (December 2025) builds on this foundation.

Watch: Whether text accuracy in generated images improves to the point where DALL-E 3’s chronic text-rendering failures are fully resolved, and whether the IP/style-emulation policy holds under legal pressure from studios and artists.