FLUX.1-dev LoRA fine-tuned on 657 reaudited Pakistani handicraft images.
Previous versions had caption/image mismatches — v3 is the clean retrain.
-
Base model: FLUX.1-dev
-
Network: LoRA dim=16, alpha=16, unet-only
-
Steps: 5000 · lr: 5e-5 · bf16 · A100 80GB
-
Dataset: 657 images, 7x repeats, corrected captions
-
Final loss: ~0.369
-
Timestep shift: 3.1582 · euler sampler
