p3-lora-bk: LoRA for BK-SDM-Small-2M (CV Hackathon 2026, P3)

(Paste this as README.md in the Hugging Face repo anu-122006/p3-lora-bk.)

• Base model: nota-ai/bk-sdm-small-2m
• Data: 1,000 COCO train2017 image–caption pairs
• Training: diffusers train_text_to_image_lora.py; UNet LoRA rank 16; 1,000 steps; batch 4; lr 1e-4; fp16; 512px; NVIDIA T4, 17 min
• Use: pipe.load_lora_weights("anu-122006/p3-lora-bk"); pipe.fuse_lora() on a BK-SDM-Small-2M pipeline
• Effect: light style adaptation toward COCO-like everyday photos; see the project's evaluation for measured results
• Licence: CreativeML OpenRAIL-M, inherited from the base model, with the same use restrictions