ezhoureal/aura_style

Flux Aura Style LoRA

This repository contains LoRA adapters for applying a radiant aura style to image edits. The style emphasizes smooth colorful gradients, ethereal haze, subtle contour lighting, and a polished cinematic glow while aiming to preserve the source subject and composition.

Training Details

• Base/edit model: black-forest-labs/FLUX.2-dev / Flux-2 Edit
• Training service: fal-ai/flux-2/lora/edit
• Steps: 1,000
• Learning rate: 5e-5
• Dataset: 25 paired image-edit examples
• Training data size: 25 target images, 25 conditioning/source images, and metadata
• Output style: radiant aura lighting, smooth colorful gradients, haze, contour highlights, cinematic glow

Default training instruction:

Apply a radiant aura lighting style with smooth colorful gradients, ethereal haze, subtle contour lighting, and a refined cinematic glow while preserving the subject and composition.

Usage

Use the adapter with compatible Image2Image inference. A useful prompt pattern is:

Transform this photorealistic image into the trained radiant aura style: smooth colorful gradients, ethereal haze, subtle contour lighting, and a refined cinematic glow. Preserve the subject identity, composition, pose, silhouette, camera framing, and important details.

Suggested starting settings:

• LoRA scale: 1.0
• Guidance scale: 2.5
• Inference steps: 28

For local Diffusers workflows, load the Diffusers-format adapter:

from diffusers import Flux2Pipeline

pipe = Flux2Pipeline.from_pretrained("diffusers/FLUX.2-dev-bnb-4bit")
pipe.load_lora_weights(
    "fal_flux2_edit_lora/pytorch_lora_weights.diffusers.safetensors",
    adapter_name="aura",
)
pipe.set_adapters(["aura"], adapter_weights=[1.0])

Evaluation Assets

eval_data/ contains the input images used for evaluation. output_sync/ contains corresponding stylized outputs generated with the uploaded adapter.

Intended Use

This LoRA is intended for stylized image editing where the source image should remain recognizable while receiving a luminous aura treatment. It works best with clear subjects, simple-to-medium complexity compositions, and prompts that explicitly preserve identity, pose, framing, and important details.

Limitations

• The adapter is style-focused and may over-apply glow or color gradients at high LoRA scales.
• The training set is compact, so unusual image domains may need prompt tuning or a lower adapter strength.
• It inherits behavior, restrictions, and access requirements from the Flux-2 Edit base model.
• The dataset and outputs are included for reproducibility and inspection.