Generating Islamic Art Style Prompts in Modern AI Image Tools
Most people trying to get quality Islamic geometric art out of image generators run into the same wall pretty quickly. The tools default to generic Middle Eastern aesthetics — wrong patterns, inaccurate arabesques, and calligraphy that looks like gibberish. I spent about three months debugging this before I figured out what actually works, and I figured I would document it. The core problem is that diffusion models have not been trained with enough rigorously accurate Islamic art references. They know the color palette, they know the general idea of arabesque and geometric tessellation, but they conflate Ottoman, Mughal, Andalusian, and Persian styles constantly. If you want a coherent output, you need to be extremely specific in your prompts, or you get a muddy mess that vaguely resembles something from a hotel lobby in Dubai.
Understanding the Advancements In Islamic Art 2 Men Bowing Prompt Style
This specific prompt format has gained traction in communities working on refined Islamic miniature and calligraphic art generation. The "2 Men Bowing" component is not arbitrary — it references a compositional tradition found in Persian and Mughal miniature painting where figures are shown in gestures of reverence or greeting, often framed within intricate architectural or floral borders. When you combine this with modern prompting techniques, you get something closer to authentic stylistic output than the usual hallucinated results. Here is how I actually construct these prompts in practice. I start with the base medium and period specification, then layer in compositional elements, pattern types, and negative prompts to steer away from common failure modes. A working prompt structure looks something like this: "Persian miniature painting, two figures bowing in courtyard setting, intricate arabesque border, fine brushwork, lapis lazuli and gold pigment palette, Safavid era style, geometric tile patterns, detailed calligraphic band, manuscript illustration, high resolution." The order matters more than most people realize. Put the style period early in the prompt, not at the end, because the model weights the beginning significantly higher.
Common Pitfalls and How to Fix Them
I ran into a persistent issue where the bowing figures kept coming out with distorted anatomy — extra limbs, fused fingers, faces that melted into the background. This happened consistently across Stable Diffusion 1.5, SDXL, and even some Midjourney iterations. The workaround was surprisingly simple: I added "anatomically correct figures, proper human proportions, distinct separate fingers" directly into the prompt, and I also enabled a face fix extension in my pipeline. That cut the failure rate from roughly 60 percent down to maybe 15 percent. Another thing nobody warns you about is the color bleeding problem. Islamic art relies on very specific, highly saturated color combinations — cobalt blue against gold, turquoise against white tile. The models love to muddy these together and produce brownish or grayish tones instead. I solved this by adding specific hex-adjacent color references to my prompts, like "vivid cobalt blue #0047AB, bright turquoise #40E0D0, metallic gold accents," which forces the model toward the right end of the color spectrum.
Get the Full Details
Tools and Setup I Actually Use
I run Stable Diffusion XL locally through ComfyUI. The workflow takes some time to set up but once configured, it is reliable. You need a good checkpoint trained on Islamic art or manuscript illustration data. The default SDXL checkpoints are no good for this — they will give you stereotypical mosque imagery every time. I fine-tuned a model on a dataset of approximately 800 high-resolution Persian and Mughal manuscript images, along with Seljuk and Ottoman tilework photographs, over about forty training hours on an A6000 GPU. If you do not have the hardware for that, there are pretrained models available on Hugging Face and Civitai that people have already fine-tuned for Islamic art generation. Look for ones specifically tagged with "miniature," "manuscript," or "arabesque." The community around this niche is small but active, and the quality of available models has improved noticeably over the past year.
Workflow That Saves Time
Here is the practical process I follow now. I write the prompt, set the resolution to 1024 by 1024 or 896 by 1152 depending on whether I want a square or portrait orientation, use a CFG scale between 4 and 6, and run about 30 to 40 steps with a DPM++ 2M Karras sampler. I generate eight images at a time and pick the best one, then use inpainting to fix any remaining issues rather than regenerating from scratch. This approach usually gets me a usable result in about fifteen minutes, compared to the two or three hours I was burning before I figured out the right settings. The main limitation I have to accept is that these models still cannot reliably render coherent Arabic calligraphy. If you need actual legible script in the artwork, you will need to generate the image without text and overlay the calligraphy separately using design software. Several people in the community have tried training models specifically for calligraphic accuracy, and while there have been incremental improvements, the results are not consistent enough for production work yet. It is something to watch but do not bet on it being solved soon.
Where to Find Resources
The checkpoints and LoRAs I reference are hosted on Hugging Face under repositories like "Islamic-Art-SDXL" and "Persian-Manuscript-FineTune." The ComfyUI workflow files for this setup are shared on GitHub. For dataset images, the Metropolitan Museum of Art and the British Museum have released high-resolution public domain scans of Persian and Mughal manuscripts that work well for fine-tuning. The Aga Khan Museum also has an open access digital collection that is useful, though their image resolution varies. If you are just starting out and do not want to deal with local installation, WebUI-based services that host SDXL checkpoints are available through various AI art platforms, though they tend to be slower and less customizable than running things locally. The quality gap between a properly tuned local setup and a generic cloud service is significant for this particular use case, so I would recommend sticking with local if you can manage it.
