network-centrality-labs/jewel-images

💍 Museum-Style Jewellery LoRA — SDXL

This model is a DreamBooth LoRA fine-tuned version of Stable Diffusion XL, trained to generate museum-quality images of historical jewellery. It learns the visual language of catalogue photography: controlled lighting, high-resolution metal textures, gemstone reflectivity, and authentic archival aesthetics.

The LoRA specialises in objects resembling those found in the Victoria & Albert Museum collection, using a synthetic token “TOK” to anchor the learned style.

✨ Model Details

• —Model type: LoRA adapter for SDXL
• —Base model: stabilityai/stable-diffusion-xl-base-1.0
• —VAE: madebyollin/sdxl-vae-fp16-fix
• —Fine-tuning method: DreamBooth LoRA
• —Training resolution: 512×512
• —Domain: Jewellery, catalogue photography
• —Task: Text-to-image generation

🎯 Intended Use

This model excels at:

• —Generating museum-style jewellery photos
• —Creating catalogue-ready rings, necklaces, bracelets
• —Rendering metals, gems, filigree, and antique forms
• —Producing historically coherent designs
• —Visualising jewellery concepts for creative or research use

Not recommended for:

• —Photorealistic human portraits
• —Non-jewellery objects
• —Safety-critical or factual applications
• —High-risk generative design for real gems/metalwork

📚 Training Data

Data was collected programmatically from the Victoria & Albert Museum’s IIIF API, querying for “Coronet” and related jewellery classifications.

Each image includes:

• —high-resolution IIIF image
• —systemNumber
• —title / summary
• —curated metadata

Synthetic training captions were produced using GPT-4.1 with detailed instructions focusing on materials, craftsmanship, gemstones, condition, and stylistic cues. This produced a consistent metadata_gpt4.jsonl file.

🔧 Training Configuration

• —Steps: 2000
• —Checkpoints: ~700 steps
• —Batch size: Small with gradient accumulation
• —Optimiser: 8-bit Adam
• —Precision: FP16
• —SNR γ: 5.0
• —Prompt token: "TOK"
• —Framework: Custom Python training pipeline

The LoRA was trained to reproduce metal shine, stone translucency, engraving, patina, and controlled archival lighting.

🧠 Model Behavior

The model produces:

• —Clean, professional product-style photos
• —Sharp metal edges and realistic gemstone reflections
• —Neutral, controlled museum-style lighting
• —Historically plausible jewellery objects
• —Minimal compression artifacts or distortions

Example Characteristics

• —Strong adherence to the “TOK” concept token
• —Good compositional control
• —No watermark artifacts (unlike the base SDXL)
• —High material fidelity even at small scales

⚠️ Limitations and Biases

• —May invent non-existent gemstones or designs
• —Limited to jewellery-like structures
• —Not trained on modern fashion photography
• —Depends strongly on correct prompt structure
• —Synthetic captions may introduce subtle biases

📊 Evaluation

Evaluation is qualitative:

• —Fine-tuned model removes watermark-like patterns
• —Strong improvement in metal realism vs base SDXL
• —Gemstone rendering is materially consistent
• —Museum-style lighting is preserved
• —Produces coherent outputs even with complex prompts

🔐 License

Released under the MIT License. Users of the LoRA must also comply with the base model’s SDXL license.

🙏 Acknowledgements

• —Image data from the Victoria & Albert Museum IIIF API
• —Captions generated with GPT-4.1
• —Base model by Stability AI
• —LoRA training enabled by PEFT + Diffusers

📣 Citation

If you use this model, please cite:

• —Stability AI (SDXL)
• —Victoria & Albert Museum (image source)
• —GPT-4.1 (caption generation)