Marcus Johansson tests ways to choose a photo’s focal point: the spot a website should keep when making different crops. His Drupal use case is concrete. Older image libraries often lack those editor-marked points, so new layouts crop everything around the centre.

Measure what the crop keeps

Johansson uses 19 Pexels photos, marks his own reference points, and repeats each method-and-prompt combination three times. He measures distance from his point and whether it survives square, wide, tall, and close-thumbnail crops.

His aggregate results across the prompts report:

  • Jev Box: mean error 8.7% of the image diagonal, reference point retained in 89% of crops, about 0.5 seconds per photo.
  • GPT-5.5 asked for coordinates: 6.2% mean error and 96% retention, about 6.8 seconds per photo.
  • Centre baseline: 14.4% mean error and 79% retention.

These are Johansson’s measurements, unreproduced here. They measure agreement with one editor’s choices, rather than a universal definition of a good crop.

Keep price settings together

Johansson estimates $0.06 per 1,000 photos for Jev and $15.81 for direct GPT-5.5 coordinates in the aggregate table. Those are model-call costs calculated from measured tokens and list prices; local models’ displayed zero excludes hardware and electricity.

His separate animation uses each model’s best prompt, with different prices. It repeats the sample photos to illustrate a larger job; it is not a fresh thousand-image evaluation.

Finding a box leaves a choice to make

  • Jev returns a box, and the workflow uses its centre as the focal point. A well-shaped box can still select the wrong subject.
  • Object detectors often supply the right candidate, but choosing which box matters remains difficult. A person’s box centre also tends toward the chest, while an editor may want the eyes.
  • GPT-5.5’s self-correction loop improves the author’s mean error only modestly, while adding calls and waiting time.
  • Johansson acknowledges that his Qwen pointing setup may be imperfect, so its weaker aggregate result needs that qualification.

Keep a review step for ambiguous groups and expensive mistakes. Typed output is not a correct decision, and keeping a single point inside a crop does not guarantee good composition.

Caveat: the full article and its tables were read; model calls, timings, costs, and Drupal integration were not reproduced.