michiyomi turns crowdsourced Mapillary street imagery into structured, queryable descriptions using vision-language models (VLMs). Search 588,220 scenes across parts of Tokyo’s 23 wards by latitude and longitude—no signup or API key required. Each photograph captures the street as it looked in its own capture year, so every record states when it was captured and which AI generation described it. The interface and documentation are available in English; authoritative scene narratives returned by the API remain in their original Japanese.
The same API powers two experiences: a map and a walking AI demo. Both call the live API available at the bottom of this page. Records of change over time (Change Map) are available as a sub-map from the Overview Map.
See the distribution of all 588,220 described scenes at a glance. Explore lenses for walkable width, marking wear, and green-view ratio; street-level views; featured places; and change-over-time records. Coverage density is shown honestly, including where observations are sparse.
Open the Overview Map →A seven-second pedestrian-view demo on a Ginza street, showing how coordinate-based descriptions can support a walking experience.
Watch the Walking Demo →The starting point is a single street photograph published on Mapillary.
The AI (VLM) describes only what is visible in the photograph, producing statements such as "Sidewalk: present on the left (separated sidewalk)." No external knowledge is injected: if something is not visible, the model cannot claim it. That constraint is the foundation of the data’s trustworthiness.
Note: Numeric values such as widths are image-based estimates with confidence scores; they are not field measurements.
AI prose that mixes facts with interpretation is not reliable enough to use. The data model enforces their separation.
Each record keeps observed facts (metadata), deterministic computation (machine), and AI interpretation (analysis) in three separate layers. The API returns them separately, allowing AI claims to be checked against the record’s own evidence.
A photograph does not show where north is. The AI reports only image-relative positions; absolute direction is computed geometrically from adjacent frame coordinates and then combined with that description. Transient details such as passers-by are excluded from the prompt from the start.
Every one of the 588,220 records includes its capture year, the AI generation that described it, and the data release. Every description can be traced back to the source photograph by image ID.
The source material is public Mapillary street photography contributed by people. Because it is available under CC BY-SA 4.0, the AI-derived data can be shared in full under the same terms. The complete 588,220-record dataset is available on Hugging Face.
Endpoint reference, MCP connection guide, and examples
Open /en/docs → Technical white paperDesign principles, including layer separation and adjudication
Open /en/docs/whitepaper → Development historyThe 38 days from kickoff on July 16 to submission on August 22
Open /en/docs/history → API Status JSONrelease 2026-08-22-r1 status (published · raw JSON)
Open /v1/meta → DatasetAll 588,220 records in Parquet (CC BY-SA 4.0)
Hugging Face ↗Enter a latitude and longitude to call the production GET /v1/coverage and GET /v1/scenes/nearby endpoints and inspect their raw JSON. No signup or API key is required. Start with the Ginza preset.
