marwantosolve/archforge-mamluk-cairo-lora

ArchForge — Mamluk / Islamic Cairo LoRA

A FLUX.1-dev LoRA adapter for Mamluk and Islamic Egyptian architecture: carved limestone facades, pointed arches, muqarnas vaulting, mashrabiya screens, ablaq striped stonework, domes and minarets of historic Cairo.

Trained as part of ArchForge — a time-boxed proof of concept, not a production model. Read the limitations below before using it.

Training

Base model	black-forest-labs/FLUX.1-dev
Method	LoRA, transformer attention layers only (to_k, to_q, to_v, to_out.0)
Base quantisation	NF4 (bitsandbytes), 4-bit
Resolution	512
Rank / alpha	8 / 8
Precision	fp16 + GradScaler
Optimizer	8-bit AdamW
Learning rate	0.0001, constant schedule, no warmup
Batch / accumulation	1 / 4
Steps	400
Hardware	Google Colab free tier, Tesla T4, 16GB

Dataset

marwantosolve/archforge-mamluk-cairo — 39 curated 512px images, per-image licence and attribution recorded.

Usage

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.float16,
).to("cuda")
pipe.load_lora_weights("marwantosolve/archforge-mamluk-cairo-lora")

image = pipe(
    "mamluk_architecture, a historic Islamic courtyard in Cairo, "
    "carved limestone facade, pointed arches, architectural photography",
    num_inference_steps=20,
    guidance_scale=3.5,
    height=512,
    width=512,
    generator=torch.Generator("cpu").manual_seed(0),
).images[0]
image.save("out.png")

The trigger token is mamluk_architecture. It is not required for the style to appear, but it is what the captions conditioned on.

Limitations

• 39-image dataset, 400 steps. This is a style demonstration, not a robust style model. Expect weak binding and inconsistent results.
• Cairo-weighted and monument-biased. Wikimedia contributors photograph well-documented, tourist-accessible buildings, and the adapter inherits that bias.
• Non-commercial. FLUX.1-dev is released under the FLUX.1-dev Non-Commercial License; this adapter inherits that restriction.
• Documented failure cases — see the evaluation report linked from https://github.com/marwantosolve/ArchForge. They are recorded deliberately rather than omitted.