chiikabu-labs/chipoint-2

chiikabu.cc  ·  discord  ·  twitter / x

Chipoint v2 is a SOTA geolocation model that is open-weight and beats the published frontier closed-weight model (Pinpoint, June 2026), its predecessor (Chipoint v1, July 2026), and all models on the OSV-5M street-view benchmark. This repo hosts both arms:

• street-view arm — the 4.89M-frame OSV-5M gallery index, the projection heads, and the 23-feature reranker pipeline (streetview_k512_joint.pt, K512, per-threshold confidence head)
• general-photo arm — the unified 14.5M photo gallery (MP14 10.9M + MP16-Pro 4.1M) and the 23-feature rerankers (photoarm_k768_thr.pt / photoarm_k768_sr.pt / photoarm_k512_thr.pt)
• router_osv_flickr.pt routes queries between the two arms for the best accuracy on a provided image.

Street-view arm

OSV-5M benchmark

OSV-5M test set (210,122 images). Acc@Xkm is the share of predictions within X km. GeoScore is 5000·exp(-d/1492.7), higher is better.

Model	@25km	@200km	@750km	@2500km	GeoScore	Mean err
OSV-5M Baseline	–	–	–	–	3361	1814 km
GeoCLIP	21.5	52.1	72.1	–	–	–
GRE	9.7	35.6	72.5	91.1	–	1192 km
LocDiff	11.0	46.3	77.0	88.2	–	–
RFM (S2S)	–	–	–	–	3767	1069 km
HierLoc	–	–	–	–	3963	861 km
Pinpoint (retrieval only)	32.1	65.6	82.8	92.8	4035	784 km
Pinpoint (full)	35.6	67.5	83.7	93.2	4114	743 km
Chipoint v1	37.63	70.56	84.94	93.36	4174	716 km
Chipoint v2 · street (full pipeline)	44.8	81.1	92.1	96.8	4485	383 km

Chipoint v2's median error is 32 km (v1: 51 km). The full-pipeline line is retrieval (K=64 pool) + the K512 23-feature reranker with a multiscale blend on the full 210,122-image test split. For reference, raw retrieval without reranking scores 34.89 / 61.96 / 74.60 / 85.34, GeoScore 3742, mean 1427 km.

General-photo arm

Running the model

Instructions on how to run Chipoint v2 locally are in the RUN_GUIDE.md file. Quick check:

python run_chipoint.py weights/streetview_k512_joint.pt
python run_chipoint.py weights/photoarm_k768_thr.pt