wangchenlong/temp

Codex-generated Blender street scenes (WebDataset)

Procedural Blender 5.0 reconstructions of 5 street-view reference photos, each written by an OpenAI codex agent (codex exec, reasoning effort xhigh) as a single scene.py.

dir	model	prompt
blender/	gpt-6.1-sol	v1 (no label requirement)
blender_astra/	gpt-6-astra	v2 (jobs/prompt_blender.md: fixed outputs, semantic label property on every object, realism)

Format

One tar shard per run (<run>/<run>-000000.tar), one sample per scene. Sample key = scene name, e.g. 03_europe_historic_facades. Members (<key>.<file>, nested paths flattened with _):

• meta.json – run, model, effort, timing, token stats, file list
• scene.py – generator script (blender -b --factory-startup -P scene.py rebuilds scene.blend + render.png)
• scene.blend, render.png, ref.jpg (4096 px reference), NOTES.md, prompt.md (v2 only)
• intermediate iteration renders / logs, codex_events.jsonl (full agent trace incl. token usage)
• viewer_objects.json (object table: name, label, parent, bbox, tris), viewer_mesh.npz (world-space triangles per label)

import webdataset as wds
ds = wds.WebDataset("hf://datasets/wangchenlong/temp/blender_astra/blender_astra-000000.tar")
for s in ds:
    print(s["__key__"], sorted(s))

Reference image sources

01–02: HoliCity panoramas (non-commercial terms); 03–05: Mapillary Vistas v2.0 (Research Use License, non-commercial). Reference images remain under their original licenses.