Rishitha05479/data266-lab2-sd15-lora-run2

DATA 266 Part 2 : LoRA Rank 32

Stable Diffusion 1.5 fine-tuned on the Naruto BLIP Captions dataset for SJSU DATA 266 Lab 2.

Experiment: run2

Selected final model: Yes

Training

Setting	Value
LoRA rank	32
LoRA alpha	32
Trainable parameters	6,377,472
Learning rate	0.0001
Steps	4000
Batch size	1
Effective batch size	4
Resolution	512 x 512
Seed	266
GPU hours	2.301
Peak GPU memory	4.47 GB

Evaluation

Metric	Value
Original KID (lower is better)	0.01256
Matched KID (12 generated images)	0.01391
CLIP score	0.3041
LPIPS diversity	0.682

The matched KID calculation uses identical unblocked in-domain prompt/seed combinations for all three models.

Limitations

The dataset is small and focused on Naruto-style imagery. Fine-tuning improves domain-specific appearance but can reduce prompt adherence, especially for out-of-domain and compositional prompts.

Some generated images were blocked by the safety checker. Blocked outputs were excluded from the image-quality metrics.

The small KID sample size limits the strength of statistical conclusions.

License and intended use

The base model is released under CreativeML OpenRAIL-M. The training images depict copyrighted characters, so this adapter is for non-commercial research and coursework only and must not be used to produce or distribute material commercially.