joshuachin/flux-hudson-river-school-style-lora

Flux Hudson River School style LoRA

<Gallery />

This is a LoRA trained on FLUX.1-dev to emulate the style of the Hudson River School art movement, known for its romantic and detailed American landscapes.

It excels at creating images with the school's signature characteristics: luminous, hazy light; a sense of epic scale and atmosphere; and a deep, spiritual connection to the landscape.

The goal of this LoRA is to provide artists with a tool to evoke the 19th-century Romantic conversation between humanity and nature, allowing the application of this grand, atmospheric style to subjects both classical and modern.

Inference Settings

You should use hrs_style to trigger the image generation.

IMPORTANT: For the best results, use an initial LoRA strength of 1.5.

• —For subjects similar to the training data (landscapes, pastoral scenes), a lower strength of 1.0 - 1.2 works well.
• —For "out-of-distribution" subjects (sci-fi, portraits, futuristic cities), a higher strength of 1.3 - 1.5 is needed to ensure the style is applied strongly and overcomes the base model's default aesthetic.

Training and Data

This LoRA was trained on the black-forest-labs/FLUX.1-dev model.

• —Dataset: 37 curated images from the Hudson River School. All images are in the public domain.
• —Captions: Sourced from the National Gallery of Art's official curatorial and visual descriptions, providing rich thematic and compositional context.
• —Training Steps: 3000
• —Learning Rate: 1e-4
• —Rank: 32 for linear layers, 16 for convolutional layers.

The full methodology, including the Python scripts for data curation, a complete list of the artworks used, and the final training configuration, is available in the GitHub repository for this project.

Download model

Download them in the Files & versions tab.