mmlkcairo-sdxl-lora

mmlkcairo — Mamluk Cairo architecture style LoRA (SDXL)

A style LoRA for SDXL that renders Mamluk-era Cairo architecture (1250–1517): weathered limestone, crenellated parapets, honeycomb (muqarnas) portal hoods, tiered minarets, carved domes on tall drums, keel arches. Trigger word: mmlkcairo.

Built in one day (23 Sep 2026) as a self-directed project: dataset collection, captioning, two training runs, and a controlled evaluation of hallucination, style proportions and style bleeding.

Recommended: v2/mmlkcairo_sdxl_v2-000002.safetensors at LoRA weight 0.8.

Rows = prompts, columns = base SDXL · v1 · v2 at weight 0.6 / 0.8 / 1.0. Rows marked NOTRIG have no trigger word (leak test).

Usage

import torch
from diffusers import StableDiffusionXLPipeline, AutoencoderKL

vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", vae=vae,
    torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.load_lora_weights("Amr292/mmlkcairo-sdxl-lora",
                       weight_name="v2/mmlkcairo_sdxl_v2-000002.safetensors", adapter_name="mmlkcairo")
pipe.set_adapters("mmlkcairo", adapter_weights=0.8)

image = pipe("mmlkcairo, a mosque facade seen from street level, tall portal in the center, minaret on the right, afternoon sun",
             negative_prompt="blurry, lowres, text, watermark, deformed",
             num_inference_steps=25, guidance_scale=7.0).images[0]

Prompt with plain object words (facade, dome, minaret, portal, courtyard, interior) plus viewpoint and light. Don't add style words — the trigger carries the style.

Dataset

Amr292/mamluk-cairo-style-dataset — 42 photos of 17 Mamluk buildings from Wikimedia Commons, each with its license and author recorded.

• Era control: buildings with major Ottoman-era parts (Cairo Citadel, Bayt al-Suhaymi, the Blue Mosque) were excluded so later styles aren't learned as Mamluk. Historical prints and postcards were excluded so sepia and grain aren't learned as "the style".
• Balance: max 5 images per building; mix of facades (11), domes (8), minarets (8), portals (6), courtyards (5), interiors (4).
• Captions describe everything except the style. Every caption starts with mmlkcairo, then names subject, viewpoint, light and all clutter (cars, wires, loudspeakers, scaffolding). Style words (Mamluk, muqarnas, ablaq, ornate, carved…) are banned and checked by script, so the style is learned by the trigger rather than by plain words, and clutter stays promptable instead of appearing by itself.

Training

kohya sd-scripts, SDXL base 1.0, single Kaggle T4 (fp16). LoRA rank 16 / alpha 16, U-Net only, AdamW8bit, lr 1e-4 cosine, min-SNR γ=5, aspect-ratio bucketing, 10 repeats.

Only one variable changed between each step, so each difference in the grids can be attributed to it.

Evaluation

Same 7 prompts × 2 seeds (42, 7) through every model, so each row differs only by the LoRA:

• Style accuracy & proportions — facade, dome, minaret, interior
• Generalization — a building type not in the data ("a public library")
• Leak tests — "a modern office building" and "a mosque facade", both without the trigger
• Clutter — the negative prompt deliberately omits "people, cars", to see whether the LoRA adds them by itself

Grids: v1 epochs · v1 vs v2 · strength sweep

Known limitations

• Little ablaq (striped two-tone masonry) — under-represented in the 42 images.
• Framing bias toward low street-level angles with building edges in frame, inherited from the photos.
• "Dome" prompts sometimes produce minaret-like towers, and at one seed a dome prompt gives an interior view; dome exteriors seen from a distance are rare in the data.
• Occasional artifacts: a fake Arabic inscription band (v1 epoch 2), floating lamp shapes in some interiors.
• Without the trigger, v2 drifts slightly from base SDXL's mosque (more teal tile) — not Mamluk, but not identical to base.
• 42 images is small; production style LoRAs typically use 100–300.

Next steps

1. Grow the dataset to ~150 images with deliberate ablaq, distant dome and courtyard coverage, and more eye-level shots.
2. Color-match regularization images to the photo set (or use real non-Mamluk photos) to remove the cause of the cast.
3. Train on Flux.1-dev and compare at matched steps.
4. Add a quantitative check (e.g. CLIP similarity to held-out Mamluk photos vs. to Ottoman/Persian references) next to the visual grids.

Files

• v2/mmlkcairo_sdxl_v2-00000{1,2}.safetensors, v2/mmlkcairo_sdxl_v2.safetensors — v2 checkpoints (ep2 recommended)
• mmlkcairo_sdxl-00000{1,2,3}.safetensors, mmlkcairo_sdxl.safetensors — v1 checkpoints
• v2/reg_images/ — the 42 regularization images and captions
• evaluation*/ — all renders and grids · notebooks/ — every notebook used, in order

Credits

Photos: Wikimedia Commons contributors — authors and licenses per image in the dataset's metadata.csv. Base model: Stability AI (SDXL 1.0). VAE: madebyollin/sdxl-vae-fp16-fix. Trainer: kohya-ss/sd-scripts. Author: Amr Mohammed · GitHub