Free Style LoRA Meta

This dataset contains metadata and demo images for LoRA (Low-Rank Adaptation) models evaluated across three base model architectures. It serves as a reference for understanding LoRA training quality, visual style/content characteristics, and evaluation configurations.

Data Structure

free_style_lora_meta/
├── flux/                          # FLUX-based LoRA evaluations
│   └── {lora_id}/
│       ├── {lora_id}.json         # Main metadata (training config, trigger words, model info)
│       ├── {lora_id}_img{1-6}.json # Per-image generation parameters
│       └── demo_images/
│           └── {lora_id}_img{1-6}.jpeg  # Generated demo images
├── illustrious/                   # Illustrious-based LoRA evaluations
│   └── {lora_id}/
│       ├── {lora_id}.json
│       ├── {lora_id}_img{1-6}.json
│       └── demo_images/
│           └── {lora_id}_img{1-6}.jpeg
└── qwen/                          # Qwen-based LoRA evaluations
    └── {lora_id}/
        ├── {lora_id}.json
        ├── {lora_id}_img{1-6}.json
        └── demo_images/
            └── {lora_id}_img{1-6}.jpeg

Distribution

File Descriptions

{lora_id}.json (Main Metadata)

Contains the primary information about each LoRA model:

• Model source and download URL
• Training trigger words / activation tokens
• Base model version and architecture
• LoRA rank, training steps, and hyperparameters
• Associated tags and categories

{lora_id}_img{N}.json (Per-Image Parameters)

Generation parameters used to produce each demo image:

• Prompt and negative prompt
• Sampling method, steps, CFG scale
• Seed and resolution

demo_images/ (Visual Demos)

Generated sample images (JPEG) that demonstrate the LoRA's effect. Typically 6 images per LoRA, showing the model's capability across different prompts.

Base Model Descriptions

• FLUX: High-quality text-to-image diffusion model known for prompt adherence and photorealistic output.
• Illustrious: A community-driven anime/illustration-focused model, excelling at stylized 2D artwork.
• Qwen: Qwen-based multimodal architecture adapted for image generation with instruction-following capabilities.

Use Cases

• LoRA Quality Evaluation: Compare generation quality across different LoRAs and base models.
• Style/Content Classification: Use metadata and demo images to build style or content classifiers.
• Triplet-based Similarity Judgment: This data supports triplet evaluation pipelines for measuring style/content similarity between LoRA outputs.
• Training Recipe Analysis: Study how different training configurations affect output quality.

License

This dataset is provided for research and evaluation purposes.