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

PhotoGuider measures four photographic tokens with compact learned scorers to select 360° view directions, rank and refine crops, output camera parameters, and generate measurement-conditioned explanations.

Traditional image cropping assumes that the photograph already contains the composition worth keeping and asks only where to crop. A full 360° scene offers no such guarantee because the viewing direction has to be discovered before anything can be framed. Existing 2D cropping methods assume a pre-shot photograph, while panoramic framing introduces an earlier direction decision.

PhotoGuider extracts 60 composition-aware features with foundation and task-specific vision models and partitions them into Subject, Composition, Aesthetics, and Integrity tokens. Candidate framings are generated from DINOv3 attention, InSPyReNet masks, and box geometry; a DPP shortlists ten, and a LightGBM ranker scores the top three using 17D core and 60D full feature tiers. Gated refinement recenters, expands, or shrinks the chosen frame when LightGBM regressors predict sufficient IoU gain. For 360° scenes, perspective views are scored by a token-grouped additive view scorer, the best direction is reprojected, and the crop is converted to pan, tilt, field of view, and aspect ratio for re-rendering. Explanations come from TreeSHAP attributions and a QLoRA-tuned MLLM conditioned on measured tokens and geometry.

• PhotoGuider full has the highest mean mIoU and lowest mean BDE averaged across FCDB and FLMS; fast mode reports average BDE 0.050 and IoU 0.797. §4.2
• In the blind 360° ranking study, PhotoGuider gets 53.2% first-place rankings versus 27.3% for Venus and 19.5% for CACNet and wins 59 of 100 scenes. §4.3
• In the double-blind explanation study, PhotoGuider explanations receive 40.0% of 250 votes, ahead of Venus at 32.8% and ArtiMuse at 27.2%. §4.2
• Fine-tuning the explainer lowers detector-measured contradiction rate from 6.31% to 1.55%. §4.2
• Full refinement improves mIoU from 0.754 to 0.759 on FCDB and from 0.820 to 0.835 on FLMS while reducing BDE. §A.6
• Token ablation shows Subject and Composition as strongest individual feature groups and all four tokens combined as best overall. §4.2

• Framing adjustments rely on discrete step fractions, so targets that fall between sampled intervals are approximated.
• View selection uses a single argmax, returning one frame even in scenes with multiple strong regions, regardless of user preference.
• The explanation preference study uses five images, leaving broader generalization of the explainer's preference open.

• PhotoGuider is a measure-then-frame method that answers 360° direction selection and framing with compact learned scorers over shared composition-aware features. §Abstract
• PhotoGuider measures 60 features on each perspective frame and groups them into Subject, Composition, Aesthetics, and Integrity tokens. §A.1
• Candidate generation creates approximately 300 boxes from eight sources; DPP selects ten framings and the top three go to the main ranker. §3.2
• The decision layer has only 63k parameters, confining the MLLM to a bounded consultant role that neither scores nor makes decisions. §1
• The ranker uses two regression-and-ranking model pairs on a 17-dimensional core set and a 60-dimensional full set to produce three quality scores. §3.3
• Recenter uses competing thirds and CLIP-attention targets; four expansion gates and a shrink fallback adjust the selected crop when predicted IoU gains exceed thresholds. §3.4
• For 360° scenes, a token-grouped additive view scorer selects a direction; the working view is cropped and converted to pan, tilt, field of view, and aspect ratio. §3.7
• Explanations use TreeSHAP attributions and a Qwen2.5-VL-7B teacher with four distillation filters; the student is trained via QLoRA. §3.6