DishaRawat/data266-team30-p2-lora-r32

Naruto-style LoRA for Stable Diffusion 1.5 (DATA 266 Lab 2, team 30) — final run lora-r32

Base model: stable-diffusion-v1-5/stable-diffusion-v1-5 (512×512). Data: lambdalabs/naruto-blip-captions, 1099 training images after removing 0 corrupted/duplicate images; 122 held-out test images (90/10 split, seed 266).

Training: LoRA rank 32 (alpha 32) on UNet attention projections ['to_k', 'to_q', 'to_v', 'to_out.0']; 6,377,472 trainable parameters; AdamW lr 0.0001 (cosine, 50 warm-up); 1000 steps, batch 2 × grad-accum 2; fp16 autocast; seed 266; 0.117 GPU-hours on NVIDIA GeForce RTX 4090, peak memory 4.05 GB.

Results (20 prompts × seeds 0/1, DPMSolver++ 30 steps, CFG 7.5):

model	KID in-domain ↓	CLIP score ↑	LPIPS diversity ↑
base	0.0936 ± 0.0130	32.28	0.816
lora-r4	0.0521 ± 0.0100	31.55	0.799
lora-r32	0.0331 ± 0.0074	31.25	0.811

Human blinded comparison vs base: see report.

Usage

from diffusers import StableDiffusionPipeline
import torch
pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
pipe.load_lora_weights("DishaRawat/data266-team30-p2-lora-r32", weight_name="pytorch_lora_weights.safetensors")
pipe("a man with spiky blond hair and an orange jacket").images[0]

Limitations: BLIP captions are generic (no character names), so specific characters cannot be requested; the adapter pulls out-of-domain and multi-character prompts toward single anime faces; training data is skewed toward male characters; KID is computed on only 20 generated images and is noisy.

License / intended use: the training images are copyrighted Naruto characters — non-commercial research and education only. CreativeML OpenRAIL-M use restrictions of the base model apply. Outputs were filtered with the SD safety checker.