Synthetic image datasets are attractive for training computer vision models, but predicting whether a model trained on them will work on real-world data usually requires expensive downstream training. The authors of a new arXiv preprint propose SADGE, a quantitative metric that estimates synthetic data utility directly from the similarity between synthetic and real image pairs, bypassing model training entirely.
Existing metrics such as PSNR, FID, and CLIP mostly measure semantic appearance, while geometric similarity is sometimes used to assess domain gaps. The authors found that neither type of metric alone reliably predicts downstream performance across object detection, semantic segmentation, and pose estimation. Instead, the strongest predictor came from a constrained bilinear fusion of DINOv3 appearance similarity and MASt3R geometric consistency.
Across five public synthetic-to-real benchmark families and 15 dataset-level variants, SADGE achieved the strongest association with transfer performance under both linear and rank-based criteria, with Pearson r=0.88 and Spearman rho=0.77. The paper, submitted to WACV 2027, argues that the non-linear interplay between appearance and geometry is what determines synthetic data utility, not either factor in isolation.