4Fever4/siwa-kershef-architecture
Siwa Kershef Vernacular Architecture — curated captioned dataset 112 curated, hand-captioned photographs of the earthen (kershef: salt-mud and rock) architecture of Siwa Oasis, Egypt — Shali fortress, Aghurmi, restored houses, the old mosque, and new buildings built in the traditional technique. Built to train a style LoRA whose trigger k3rshef carries the material/massing vocabulary of the style without bleeding into other architectural styles. How it was built… See the full description on the dataset page: https://huggingface.co/datasets/4Fever4/siwa-kershef-architecture.
Siwa Kershef Vernacular Architecture — curated captioned dataset
112 curated, hand-captioned photographs of the earthen (kershef: salt-mud and rock) architecture of Siwa Oasis, Egypt — Shali fortress, Aghurmi, restored houses, the old mosque, and new buildings built in the traditional technique. Built to train a style LoRA whose trigger k3rshef carries the material/massing vocabulary of the style without bleeding into other architectural styles.
How it was built
Balance decision: Commons is dominated by photographs of the Shali ruins. Left unchecked, a LoRA learns "decay" as the style. Ruins were capped to ~45 % and every image is captioned with its condition (restored, partially restored, newly built, eroded ruin) so that condition is separable from style.
Captions
Hand-written with a controlled vocabulary — see CAPTIONING.md. Format:
k3rshef, <view>, <subject>, <elements...>, <condition>, <context>, <light>, photoElement words (palm-trunk beams, thick tapered columns, small window openings, external staircase, crenellated parapet …) are captioned so they can be requested or omitted at inference. Material, texture and colour are intentionally not captioned so they bind to the trigger.
Licensing & attribution
Each image keeps its original licence. train/metadata.csv lists, per image: source page, author, licence and the caption. Images: 57 Public Domain, 33 CC BY-SA 4.0, 10 CC BY 4.0, 7 CC BY 3.0, 5 CC BY 2.0. Captions and curation: CC BY-SA 4.0.
reg/ — regularisation images (used by the v2 experiment)
48 images generated with base SDXL (not real photos) of neighbouring styles — Nubian, modern Cairo, Mamluk, Ottoman, Mediterranean, riad, desert resort, etc. — captioned without the trigger, used for prior preservation in the v2 LoRA. See the model repo for why v2 over-regularised and v1 at strength 0.6 is the recommended result.
