ChatGPT is undoubtedly the most well-known conversational agent among the general public. The chatbot owes this reputation to its many qualities in various fields, but image generation wasn't really one of them. This is a problem for OpenAI, which must face increasingly strong competition, such as Imagen 3 (Google), Aurora (xAI and Grok), or specialized tools (Midjourney, Leonardo, Firefly, etc.). The American firm may have found a solution to prevent its 400 million users from looking elsewhere.
OpenAI has just unveiled a new image creation tool integrated directly into ChatGPT. This new model, called GPT-4o Image Generation, replaces the old Dall-E model and should help the company catch up in this area.
The strengths of the new tool
While the portmanteau Dall-E is funny and easy to remember, the company is not joking with this new tool. It doesn't really have a name and is not an instinctive model, but an extension of the "omnimodal" GPT-4o model that appeared last year. It integrates natively with ChatGPT, and this constitutes a first important advantage for users. GPT-4o Image Generation aims to make Dall-E forgotten by distinguishing itself through its ability to produce highly realistic images, paying particular attention to detail. One of the strengths of this new model is its ability to efficiently generate text in images, filling a significant gap in previous tools. It also displays better comprehension and does not require the user to know how to write prompts.
This model can process between 10 and 20 requests in a single query, allowing users to describe very detailed scenes. OpenAI also highlights the versatility of GPT-4o Image Generation and the tool proves, once again, impressive in the exercise. It can generate a wide variety of formats, from realistic photographs to comics, infographics, diagrams, and promotional visuals for social media.
Rather than simply creating aesthetic or surreal images, OpenAI focuses on producing "useful" images:
On social media, the ability to edit an image in the "Ghibli style" is causing a stir:
Furthermore, the model has the ability to gradually modify an existing image, and users can request adjustments, additions, or transformations to a generated image. A solution that puts ChatGPT in direct competition with photo editing solutions such as Photoshop.
OpenAI acknowledges the model isn't perfect
ChatGPT's new image generator doesn't hold back on generating celebrities, but OpenAI acknowledges it still has room for improvement. Current limitations include sometimes overly tight cropping of long images, occasional hallucinations (fabricated information), difficulty with highly accurate rendering of complex concepts (such as a complete periodic table) or text in non-Latin languages, and editing accuracy that still needs improvement.
Furthermore, the company says security "remains a priority" and that generated images include C2PA metadata to indicate their origin. Additionally, dangerous or policy-violating content is blocked, with tighter restrictions for images of real people.
How to try ChatGPT's new image generator?
OpenAI has chosen to make this new feature accessible to all ChatGPT users, including those using the free version. Simply access ChatGPT and ask it to generate an image, a photograph, or a drawing on the theme you want. To take full advantage of its capabilities, we recommend being as precise as possible to obtain a result that meets your expectations. Rendering times can be up to one minute.
Note that the rollout is gradual and should be completed within the next few weeks, which explains why ChatGPT may still use Dall-E in its responses.
0 Comments