---
license: mit
tags:
- stable-diffusion
- lora
- flux
- moroccan
- ghibli
- art
- architecture
- cultural
---

# Moroccan Ghibli Style LoRA for FLUX.1

This LoRA model was trained to generate Moroccan architecture and scenes in the style of Studio Ghibli. It combines the intricate details of Moroccan architecture and culture with the magical, whimsical aesthetic of Ghibli animations.

## Model Details

- **Base Model:** [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)
- **Training Dataset:** [atlasia/Ghibli-style-morocco-dataset](https://huggingface.co/datasets/atlasia/Ghibli-style-morocco-dataset)
- **Training Steps:** Multiple checkpoints available (starting from 1250)
- **Resolution:** Trained on multiple resolutions (512x512, 768x768, 1024x1024)
- **Trigger Word:** "Moroccan Ghibli studio style"

## Usage

```python
import torch
from diffusers import FluxPipeline
from huggingface_hub import hf_hub_download

# Load base model
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", 
    torch_dtype=torch.bfloat16
)

# Load LoRA weights (choose your preferred step)
lora_path = hf_hub_download(
    repo_id="atlasia/moroccan-ghibli-flux-lora",
    filename="moroccan_ghibli_flux_lora_000001250.safetensors",
    revision="step_1250"
)
pipe.load_lora_weights(lora_path)

# Optional: Enable CPU offload for lower VRAM usage
# pipe.enable_model_cpu_offload()

# Generate image
prompt = "Moroccan Ghibli studio style traditional riad courtyard"
image = pipe(
    prompt,
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=50
).images[0]
```

## Example Prompts

Always include the trigger word "Moroccan Ghibli studio style" in your prompts:

- "Moroccan Ghibli studio style of a traditional riad courtyard with a fountain"
- "Moroccan Ghibli studio style of a medina street scene at sunset"
- "Moroccan Ghibli studio style of a traditional tea ceremony"
- "Moroccan Ghibli studio style of an ancient palace interior"
- "Moroccan Ghibli studio style of a garden with traditional zellige tiles"

## Training Details

- **Learning Rate:** 4e-4
- **Optimizer:** AdamW
- **Batch Size:** 1
- **Training Steps:** Starting from 1250
- **Resolution:** Multi-resolution training (512, 768, 1024)
- **LoRA Rank:** 16

## Model Branches

Each branch (`step_X`) contains a checkpoint from different points in the training, starting at step 1250:
- `step_1250`: Initial checkpoint
- `step_1500`: Mid-training checkpoint
- `step_1750`: Later checkpoint
- `step_2000`: Final checkpoint
(Actual available steps may vary)

## Important Notes

- Always use the trigger word "Moroccan Ghibli studio style" in your prompts
- The model was trained with multi-resolution support
- Best results are achieved when following the example prompt structure

## Limitations

- Works best with architectural and scenic prompts
- Requires the trigger word for optimal results
- Inherits FLUX.1 base model limitations

## Examples

[Add some example images here with their prompts]

## License

This model is released under the MIT license. However, please note that the base model (FLUX.1) has its own license and usage restrictions that must be followed.

## Acknowledgments

- Base model by [Black Forest Labs](https://huggingface.co/black-forest-labs)
- Training infrastructure using [AI Toolkit](https://github.com/ostris/ai-toolkit)