PhotoGuider: Composition-Aware Cropping and 360^{circ} View Framing with Explainable Photographic Assessments

PhotoGuider. Composition-Aware Cropping and 360^{circ} View Framing with Explainable Photographic Assessments. By Yuan-Kang Lee and three others, from MediaTek, National Taiwan University, and one other institution. Published on the second of November.

Ordinary image cropping rests on an assumption: the photograph already holds a good composition, and the only question is where to cut. The photographer has done the hard part by pointing the camera; the crop just tidies the edges. A full three-hundred-sixty-degree scene makes no such promise. It surrounds the viewer in every direction, and most of those directions are not a picture. Before anything can be framed, the viewing direction has to be discovered. Cropping alone cannot answer that.

PhotoGuider calls itself a measure-then-frame method, and it answers both questions, where to look and how to frame, with compact learned scorers. Those scorers all work from one shared set of composition-aware features. Measuring comes first, and every framing decision follows from the measurements.

The features are extracted with foundation models and task-specific vision models, then organised into four photographic tokens: Subject, Composition, Aesthetics, and Integrity. On top of that sits an explainer, a multimodal language model fine-tuned to produce explanations about composition, conditioned on the image and on the measured token assessments. So the system does not just choose a frame; it can say why, in terms of what it measured. To the authors' knowledge it is the first method to bring together explicit composition-based direction selection in panoramas, refined framing, camera parameters as output so the view can be re-rendered natively, and explanations conditioned on measurements.

On two standard 2D cropping benchmarks, PhotoGuider reaches the highest average mIoU against ground-truth crops. For panoramas there was a blind ranking study. One hundred and twenty participants ranked three cropping methods across one hundred panoramic scenes from the public Poly Haven HDRI library.

They placed PhotoGuider's framings first in fifty-three point two percent of cases. The two strong alternatives were placed first in twenty-seven point three percent and nineteen point five percent of cases.

The shift here is from cutting to choosing. In a panorama there is no picture yet, only a scene, and PhotoGuider treats finding the picture as a measurement problem. The same measurements drive the choice of direction, the final frame, and the explanation, so the reasoning it puts into words is the reasoning it actually used to frame the shot.