# Façade label source attribution

Public source ID: `pix2pix-facades-testb-1`. Original dataset reference: `facades/test/1.jpg`, right half (domain B), 256×256 RGB. The PNG is a lossless crop/conversion of the JPEG-decoded label pixels; no palette remapping was applied.

Underlying dataset: CMP Facades, Radim Tyleček and Radim Šára, *Spatial Pattern Templates for Recognition of Objects with Regular Structure*, GCPR 2013. The [official CMP dataset page](https://cmp.felk.cvut.cz/~tylecr1/facade/) explicitly identifies its license as **CC BY-SA (Attribution-ShareAlike)**. The page links to a generic Creative Commons licenses page rather than a numbered license version; this package does not invent a version.

Prepared paired dataset: Phillip Isola et al., *Image-to-Image Translation with Conditional Adversarial Networks*, CVPR 2017; Berkeley pix2pix facades.tar.gz. Recovered from [the archived publisher URL](https://web.archive.org/web/20251204204648id_/http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/facades.tar.gz). Archive SHA-256 `2f6b361ffb980763db3a7462bbe271c3a0f2ce6d8643f8534024e56681cf2f2a`, 30,168,306 bytes. The crop was independently checked pixel-for-pixel against that archive and its twelve most common quantized palette colors, recorded in `../evidence/source-proof.json`.

Packaged PNG: `facades-testB-1.png`, 70,229 bytes, SHA-256 `bb008b2e9d1a9a57e47be87512a693b100d6167c8d054dd6a64b31f1e999f55d`. Retain this attribution and ShareAlike notice when redistributing the source crop. No synthetic label map is included in the serving source allowlist.

## Editable example layouts

The editor additionally includes right-half label crops from `facades/test/36.jpg` (Window grid), `facades/test/50.jpg` (Balconies), and `facades/test/10.jpg` (Shopfront), from the same verified archive and under the same CMP CC BY-SA terms above. These PNGs preserve every decoded source label pixel; no palette remapping is applied to the packaged examples. `editor-examples.json` records member, pixel and PNG hashes. Edits and uploaded derivatives made from these dataset examples retain this attribution and ShareAlike notice.

The drawing palette reproduces the named colors from Christopher Hesse’s [original pix2pix facade editor](https://github.com/affinelayer/pix2pix-tensorflow/blob/master/server/static/index.html). The training label JPEGs have compression variation; the editor palette is a discrete drawing convention, not a claim about every training pixel. Editor uploads are normalized to this palette separately from the original stored dataset examples.