run1: Stable Diffusion 1.5 LoRA (rank 4) for naruto-blip-captions style
DATA 266 Lab 2, Part 2, team 35. Comparison run. The final model is kruthisham/data266-lab2-team35-p2-run2 (rank 32), chosen by the stated rule: lowest in-domain KID.
License: non-commercial research only. The training images show copyrighted characters, and the dataset card of lambdalabs/naruto-blip-captions declares no license.
Training: LoRA rank 4 on UNet attention layers, 797,184 trainable parameters (0.093%), lr 0.0001, 3000 steps, batch 4, 6.4 min on NVIDIA GeForce RTX 4090 (peak 9.42 GB, capped at 15 GiB). The data, steps and settings are the same as run2; only the rank differs.
Results (20 prompts × 2 seeds, DPM-Solver++ 30 steps, guidance 7.5):
| base | run1 (this) | run2 (final) | |
|---|---|---|---|
| KID in-domain ↓ | 0.1256 ± 0.0125 | 0.0240 ± 0.0095 | 0.0053 ± 0.0043 |
| CLIP score ↑ | 26.36 | 24.90 | 25.00 |
| LPIPS diversity ↑ | 0.793 | 0.776 | 0.746 |
Safety: the Stable Diffusion safety checker flagged 3 of 40 images from this run; 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-run1")
Model tree for kruthisham/data266-lab2-team35-p2-run1
Base model
stable-diffusion-v1-5/stable-diffusion-v1-5