A 2025 Comparison of AI Image Generation and Editing

In March 2025, I compared OpenAI’s 4o image generation with Reve, Imagen 3, Midjourney and another unreleased model. These are informal examples and subjective judgments from that date, not a controlled benchmark or a current ranking.

I was struck by how closely the new model followed prompts, including a request for six people of different ethnicities doing jazz hands. Some details were imperfect, and the model still had difficulties with reflections.

One deliberately difficult prompt asked for three apples on a blue, three-legged elephant near five willow trees in El Jem, Tunisia. The OpenAI examples captured the blue elephant, apples and recognizable setting, but not the requested three legs; tree counts were debatable.

In the samples I show, Imagen 3 also missed several details. Reve produced appealing images and a strong sense of place, though some object counts and shadows were imperfect. Another unreleased model generated interesting images that did not fully match the request.

For the idiom “hold your horses,” I wanted both a visual and the meaning of slowing down. I judged the 4o examples more successful at conveying the phrase and rendering text than the other examples. The sample size and my interpretation limit that comparison.

I also asked the model to turn an existing thumbnail into a three-dimensional image. The outputs preserved much of the text and drew on the original design, though details such as a logo were imperfect. I tried a similar workflow for another video thumbnail.

A four-panel human-life prompt produced labeled stages, sometimes with spelling errors. I then asked for glasses to be added to each character; the edited output largely preserved the original composition. The other models’ examples varied in how clearly they represented all four stages.

I briefly compare native image editing with Gemini in Google AI Studio. In my examples, both models can edit an image, but the results differ in how well they follow a particular instruction. This is a narrow demonstration rather than a comprehensive evaluation.

The model refused some requests but allowed an image of public figures arm-wrestling, with imperfect likenesses. I invite discussion of filtering and point to security testing of models, including attacks involving images.

I say Sora can turn images into video, but I do not recommend the examples as lifelike. I revisit the six-person hand gesture prompt and note that other samples struggled with the requested gesture or declined the request.

For a Where’s Waldo-style puzzle, the models often generated attractive crowded scenes but failed to include the object named in the caption. One Imagen example did include a recognizable target, although the text and difficulty were imperfect. This test exposes a gap between plausible imagery and a solvable task.

My overall impression in 2025 was that 4o had made a noticeable advance in prompt adherence and text rendering, even though its outputs still failed some logical constraints. These results reflect selected prompts and should not be generalized to every use or later model version.