HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
01 / 15 · 00:00

60% of guesses landed less than 1 mile away. No GPS.

INPUT1 ordinary photo

metadata removed

OUTPUT<1 mile

in 60% of 50 US photos (o3)

از یک عکس معمولی، بدون GPS، تا همان خیابان

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
02 / 15 · 00:45

What does one ordinary photo give away?

plant potair coolerbrick coloursun anglebuilding oppositesky

Guess first: on photos from Iran, how far off is the best model, at the median? 1 km · 10 km · 100 km · 1,000 km

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
03 / 15 · 01:45
LUO ET AL. · 50 US PHOTOS · o360% < 1 mile

84% within 5 miles. Photos came from Google Maps, so may have been seen in training.

BELLINGCAT · OCT 2026 · 52 NEVER-PUBLISHED PHOTOS>50% < 25 km

Top models. Good at the area, rarely exact, sometimes confidently wrong.

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
04 / 15 · 02:50
$ python3 artifact/geo_lab.py strip ~/Pictures/geo-iran --set iran EXIF Make: phone DateTime: 2026:10:… GPS 35.70 N, 51.40 E → saved to private truth.csv, removed from the copy 10 photos -> private/geo-lab/clean

The model only ever sees the clean copy. The pixels still talk.

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
05 / 15 · 03:50
6 / 6games won by PIGEON against a professional GeoGuessr player

Blind test, 458 matches: beat the Champion Division (top 0.01% of players). Median error on Street View: 44.4 km.

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
06 / 15 · 04:55
1 · PIXELSthe clean photo
2 · EMBEDDINGa row of numbers; similar photos sit close
3 · NEAREST PLACESwhich parts of the world look like this?
4 · REASONINGdrives on the right · red soil · sign font
5 · ONE POINTlat, lon

GeoRC (ACL 2026): big models rival human experts at where, still lag at explaining why.

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
07 / 15 · 05:55

Iran-Places: well-known places in Iran, ~17%.

GPT-4o16.44%
Claude 3.5 Sonnet17.07%
Llama 3.2 90B16.43%

500+ photos · ≥7 per province · relaxed exact match (province 0.5, both 1) · 2024 models · landmarks, not streets

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
08 / 15 · 07:00

20 photos · EXIF removed · 3 models + 1 human · same prompt

≤ 1 km street≤ 25 km city≤ 200 km region≤ 750 km countryreport the median, not the mean
HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
09 / 15 · 08:00
ABROAD · 10 PHOTOS[RUN] km median

[RUN] of 10 within 25 km

IRAN · 10 PHOTOS[RUN] km median

[RUN] of 10 within 25 km

ME · HUMAN[RUN] km median

all 20 photos

[RUN] fill from private/geo-lab/results.md · Homodyne test · 20 photos · Nov 2026

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
10 / 15 · 09:10
L0 original[RUN]
L1 text hidden[RUN]
L2 landmarks hidden[RUN]
L3 tight crop[RUN]

Clues are redundant: hide one, three more remain. (Cleveland: found again from waterfront and rail tracks.)

[RUN] median error, km, best model · bar widths set after the run

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
11 / 15 · 10:15

More photos online → easier to place.

LESS SEEN

More privacy today.

BUT

Less help when a flood or earthquake video needs locating.

AND

Every photo we post fills the album for the next model.

[VERIFY] Google Street View coverage in Iran

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
12 / 15 · 11:15
1Location off in the camera
2Strip metadata before sending
3Never home + street in one photo
4Hide text: signs, plates, shop names
5Post later

None of these is a guarantee. They only make it harder.

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
13 / 15 · 12:10
THIS SHOWSMetadata removal is not enough. Pixels carry location. Hiding clues helps, not fully.
THIS DOES NOT SHOWAny model's general accuracy: 20 photos, one person, one run. Indoor photos are much harder. The 60% is US-only, one model.
HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
14 / 15 · 12:50

Often the city. Sometimes the street.

$ python3 artifact/geo_lab.py score iran median [RUN] km

HOMODYNE FORECAST LEDGER · Nov 2027 retest of the same 10 Iran photos · best-model median error at least halves · p = 0.60

HOMODYNE BENCHMARK LAB · BRIDGE / HMD-B07
15 / 15 · 13:35

Next: one word exposed how much of science ChatGPT writes.

×47"delves" in PubMed abstracts, 2022 → 2024 · then it fell