LiushenbeiYH/extended_env_1016

extended_env_hf_1027

The 2026-10-27 release of extended_env_1016, prepared by the Jacobs lab, WashU. It extends the SetVL street-view node graph with indoor nodes behind real storefronts:

• 9,092 scenes in Chicago, London, Rio, Sydney, Taipei and Johannesburg, each with at least 6 street nodes;
• 8,021 indoor nodes in 3,083 scenes, connected by 8,558 entrance edges to the street nodes that see their door;
• one generated 360° indoor panorama per indoor node (5760x2880): 4,095 made with Qwen-Image 2.1 + pano360 LoRA, 3,926 with DiT360.

What changed since extended_env_hf_1025 (2026-10-25): every indoor node's type was checked by eye against its storefront (PIPELINE.md C6). 198 nodes that were a different kind of place than their type (e.g. an estate agent typed bank, a furniture showroom typed bar) got the right type and a new panorama (Qwen-Image 2.1 + pano360 LoRA); 180 nodes that are no target place (vacant units for rent, building lobbies, churches, claw-machine arcades, scrap dealers, ...) were removed with their entrance edges. What changed per node is listed in qc/type_labels_20261026/ inside the package.

Inside the package:

• README.md: folder layout, file formats, how to read the graph, and the viewer;
• PIPELINE.md: how it was made (storefront detection and acceptance, panorama generation, quality control and fixes).

Files

Every zip has the top folder extended_env_hf_1027/ and can be extracted on its own. Extracting all of them into one folder gives the complete package.

file	content	size
extended_env_hf_1027_core.zip	graph (graph_extended.sqlite), results, docs, code, viewer, OSM inputs; for every indoor node its records (entrance, placement, prompts, generation metadata) and small images (storefront crop, street view, satellite overlay)	see manifest.json
extended_env_hf_1027_panorama_5760_<City>_<k>of<n>.zip	the 5760x2880 indoor panoramas of one city (indoor_nodes/<scene>/IN<k>/pano_<method>/panorama_5760.png)	~20 GB each
manifest.json, SHA256SUMS	the file list of every zip, sizes and checksums

Only the core zip is needed for the graph and the node records. Add the panorama zips of the cities you need.

hf download LiushenbeiYH/extended_env_1016 --repo-type dataset --include "extended_env_hf_1027_core.zip" --local-dir .
hf download LiushenbeiYH/extended_env_1016 --repo-type dataset --include "extended_env_hf_1027_panorama_5760_Chicago_*" --local-dir .
sha256sum -c SHA256SUMS --ignore-missing
for z in extended_env_hf_1027_*.zip; do unzip -q -n "$z"; done

Not included in this release, to keep it smaller:

• the 2048x1024 versions of the panoramas (pano_<method>/panorama.png). The panorama column of the entrance table and the results/*.jsonl files still name them; use panorama_5760 instead.
• the intermediate images of the Qwen-Image + LoRA method (room2.png, shop.png, lora_raw.png).

The viewer in the package shows the 5760 panoramas, so it works with this release as it is.

The previous releases are in the history of this repository: extended_env_hf_1025 (12 zips, 2026-10-25, before the type check) and extended_env_1016.zip (one zip, without panoramas, 2026-10-16).

Quick start

import sqlite3
from PIL import Image
PKG = 'extended_env_hf_1027'
db = sqlite3.connect(f'file:{PKG}/graph_extended.sqlite?mode=ro', uri=True)
sid = 'Chicago:Chicago_2023_-87.62960701413965_41.90332947982964'
for row in db.execute('select indoor_label, outdoor_label, entrance_heading_deg, indoor_type, sign_text, panorama_method, panorama_5760 '
                      'from entrance where scene_id = ? order by indoor_label', (sid,)):
    print(row)
pano = Image.open(f'{PKG}/' + db.execute("select panorama_5760 from entrance where scene_id = ? and indoor_label = 'IN4'", (sid,)).fetchone()[0])

Conventions:

• Headings are bearings clockwise from north.
• The centre of an indoor panorama faces into the room; its left and right edges meet behind the viewer, at the shopfront.
• Rule for agents: enter an indoor node only if the agent's 100° view contains its entrance heading.

Not included, needed for the street views

The raw SetVL-480K street views and satellite tiles and the 360° street panoramas are not redistributed here. The node graph is that of MarcoXL/node-graph-viewer.

License and intended use

Research use only.

• The panoramas were generated with FLUX.1-dev (FLUX.1 [dev] non-commercial license) + DiT360 LoRA (MIT), Qwen-Image 2.1 (Qwen Research License) + pano360 LoRA, and HYPIR (HYPIR license). Their terms apply to the generated images.
• OpenStreetMap data in inputs/ is © OpenStreetMap contributors, ODbL.
• The storefront crops, street-view windows and satellite overlays are derived from SetVL street-view data.