sd15-naruto-lora-run2

LoRA fine-tune of Stable Diffusion 1.5 on the Naruto BLIP Captions dataset (lambdalabs/naruto-blip-captions), for DATA 266 Lab 2, Team 42. This is the run selected as final.

• Base model: stable-diffusion-v1-5/stable-diffusion-v1-5
• Dataset: lambdalabs/naruto-blip-captions (non-commercial research use; Naruto is a copyrighted character)
• LoRA rank: 16
• Training steps: 1200
• Learning rate: 1e-4

Evaluation results (selected as final run)

Metric	Base SD1.5	This run
KID (lower is better)	0.119	0.024
CLIP score	0.312	0.304
LPIPS seed diversity	0.687	0.660

Selected over the other LoRA run (rank 4) for its much lower KID (0.024 vs 0.050) against the real test set, at essentially the same CLIP score.

Limitations

Trained on a small (~1.2k image), narrow-domain dataset. Out-of-domain prompts may be ignored or produce stylistically inconsistent results. For research/educational use only.

Safety

No NSFW filter is applied by this adapter itself; outputs should be generated through a pipeline with Stable Diffusion's built-in safety checker enabled for any externally-facing use. A memorization check (closest training image by CLIP similarity, on 10 in-domain generations) found similarities of 0.90-0.92 for the closest pairs -- notably similar but short of near-exact duplication.