sdan/geospot-base

GeoSpot Base

A geolocation model built on SigLIP2-so400m (512px) that predicts GPS coordinates from images.

Model Details

• Backbone: google/siglip2-so400m-patch16-512 (frozen)
• Image Resolution: 512x512
• Embedding Dim: 512
• Training Steps: 206k
• Training Data: ~10.6M streetview images

Architecture

GeoCLIP-style contrastive learning between:

• Image Encoder: SigLIP2 vision tower + MLP projection (1152 → 512)
• Location Encoder: Multi-scale RFF encoding with learnable capsules

Usage

from geoclip.model.GeoCLIP import GeoCLIP
import torch

model = GeoCLIP(from_pretrained=False, encoder_name="siglip2")
state_dict = torch.load("model.safetensors")
model.load_state_dict(state_dict)

# Predict location from image
top_gps, top_probs = model.predict("image.jpg", top_k=5)