kruthisham/data266-lab2-team35-p2-run2

run2: Stable Diffusion 1.5 LoRA (rank 32) for naruto-blip-captions style

DATA 266 Lab 2, Part 2, team 35.

License: non-commercial research only. The training images show copyrighted characters, and the dataset card of lambdalabs/naruto-blip-captions declares no license.

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

Training: LoRA rank 32 on UNet attention layers (to_q, to_k, to_v, to_out), 6,377,472 trainable parameters, lr 0.0001, 3000 steps, batch 4, 6.3 min on NVIDIA GeForce RTX 4090 (peak 9.55 GB, capped at 15 GiB to match a 16 GB GPU).

Results (20 prompts × 2 seeds, DPM-Solver++ 30 steps, guidance 7.5): KID 0.0053 ± 0.0043 · CLIP score 25.00 · LPIPS diversity 0.746. Base model for comparison: KID 0.1256 ± 0.0125, CLIP 26.36, LPIPS 0.793. Human evaluation vs base (25 blinded pairs): win rate (style) 96% (75 votes, 3 raters); win rate (prompt) 41% (75 votes, 3 raters).

Limitations: follows out-of-domain prompts less well than the base model, inherits BLIP caption errors and character imbalance, and KID is noisy: it uses only 15 of the 20 in-domain images (NSFW-flagged ones left out) against 122 test images. The Stable Diffusion safety checker is the NSFW filter: 6 of 40 images from this run were flagged; they are published as black placeholders and left out of every metric.

Usage:

pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
pipe.load_lora_weights("kruthisham/data266-lab2-team35-p2-run2")