The dataset.
Every cut, stone, metal, and setting was rendered under matched studio conditions, deduplicated, and checked, 1,267 images across 54 attribute classes.
Method
The method, on record.
A controlled study, not a scrape. The training corpus was generated and filtered in house, then used to fine-tune SDXL with LoRA so the model holds one house style across every attribute combination.
SDXL 1.0LoRA54 classescheckpoint-1000
SDXL 1.0
Base model
LoRA
Adapter
1,267
Curated images
54
Attribute classes
30steps
Inference
7.5cfg
Guidance scale
checkpoint-1000
Deployed checkpoint
Base · SDXL 1.0Adapter · LoRADataset · 1,267Classes · 54Inference · 30 steps, 7.5Deployed · checkpoint-1000Hyperparameters · [hyperparameters: owner to confirm]
Method
The method, on record.
The fine-tune.
A LoRA adapter was trained on the clean corpus, teaching the base model the house look without retraining it end to end.
The inference.
Each concept is produced in 30 steps at guidance 7.5, with negative prompts holding the composition to a single ring on a clean studio background.
Every image here is an AI-generated design concept, not a photograph of physical stock.