The LoRA is intended to make FLUX render Mythograph Atelier prompts in a soft, abstract, worn, emotionally legible visual language.
Training Summary
Dataset: flux_lora_dataset, prepared as image/text-caption pairs
Images used: 38
Training steps: 1,800
LoRA rank/alpha: linear 128/64, conv 64/32
Optimizer: AdamW 8-bit
Precision: bf16
Training-time transformer quantization was enabled only to reduce GPU memory use; the uploaded artifact is the LoRA checkpoint, not GGUF and not a quantized model export.
