latent-inspector
🔬
How self-supervised vision models carve the same image.
Public front door for latent-inspector: live Space, fingerprints dataset, and the ONNX encoders it probes.
How self-supervised vision models carve the same image.
Note Live browser entry point: narrative deck + self-contained four-model compare report (DINOv2, I-JEPA, V-JEPA 2, EUPE) on a single elephant image.
Note Numeric evidence layer: per-model representation-geometry fingerprints + pairwise CKA and k-NN overlap for the same four models. CSV + raw compare.json.
Note Flagship model in the stack: V-JEPA 2 ViT-L/16 with image-native ONNX I/O (no frame duplication required).