OpenAI introduces Images 2.6, an update to its image generator that adds animated GIF creation and brings the annotation feature already present in ChatGPT to the API. Subjected to five practical tests covering generation, editing, and animation, the model succeeded at some tasks and left others incomplete.
- OpenAI lance Images 2.6, avec création de GIF animés et annotations dans l’API.
- L’éditeur affirme une génération 50 % plus rapide, mais ce chiffre n’est pas vérifié.
- Les tests réussissent en création et retouche, mais des erreurs persistent dans le texte.
- Les résultats restent imparfaits, donc chaque image doit être relue.
New Features Focused on Creation
Images 2.6 allows for the creation of animated GIFs, meaning short animated images in the widely used web format. It also gives developers the ability to adjust the complexity of the planning phase, a step that precedes image generation. For the user, the method remains similar to the previous version: describe the expected result, attach a reference image if needed, then specify what should remain unchanged.
OpenAI claims the update generates images 50% faster. This is a statement from the publisher, which the five tests described below did not allow us to verify through a controlled speed comparison.
Five Tests, Mixed Results
The first test requested a cinematic photograph of a glass greenhouse on a snowy ridge at dusk. Ice, condensation, tropical plants, and warm light had to coexist credibly. The author deemed the result very successful.
The second test involved a café menu. The text, uploaded to ChatGPT as a one-page PDF, contained nine lines, five prices, and the Hindi word “कॉफी”. The author considers that the obtained visual meets all the criteria, but the test’s FAQ indicates that the model did not faithfully reproduce all characters: the words and prices are preserved, but five separator dashes were omitted. The test also does not allow for a conclusion of native PDF support by the image model, as reading the document falls under ChatGPT’s functionality.
The third test involved retouching a room photograph. Using the annotation tool, a red rectangle was drawn around an armchair, whose fabric was to be changed to forest green velvet, without modifying the rest of the room. The author describes the modification as surgical.
The fourth test requested a field guide-style plate on the life cycle of the monarch butterfly, with four stages, four exact labels, and arrows indicating the sequence. The result was described as a clear layout, usable for illustrating a metamorphosis.
The final test aimed for a looping GIF depicting a copper teapot in the rain, with a swirl of steam briefly forming a heart. The result was judged to be smooth. According to the FAQ, it is a 12-frame GIF with a 2.16-second loop.
Limitations to Watch
The author points out that the results are not perfect. Errors persist in text, object outlines, and visual accuracy, which requires proofreading each final image. They conclude that Images 2.6 appears promising for creative exploration as well as practical retouching, provided the result is checked.
