lora_r8: LORA rank 8 on Stable Diffusion 1.5 (Naruto)

Selected final run. DATA 266 Lab 2, Part 2, Pair_Programming_Team_03 (Yash).

Base model: stable-diffusion-v1-5/stable-diffusion-v1-5. Adapter on UNet attention (q, k, v, out) and feed-forward layers, rank 8, alpha 8.
Data: lambda/naruto-blip-captions, 1099 training images after removing corrupted and duplicate images (90/10 split, seed 266), 512x512.
Training: 2000 steps, batch 4 (7.28 epochs), AdamW lr 0.0001, cosine schedule, Min-SNR gamma 5, torch.bfloat16 mixed precision; 2,992,128 trainable parameters; 973 s on NVIDIA RTX PRO 4500 Blackwell (0.2703 GPU-hours), peak memory 4.74 GB.
Results (20 fixed prompts x 2 seeds, DPM-Solver++ 30 steps, guidance 7.5): KID (in-domain vs test) 0.0243 (base 0.1393); CLIP score 30.68 (base 31.91); LPIPS seed diversity 0.772 (base 0.794); NSFW flags 2/40.
Limitations: generic BLIP captions, narrow subject range (anime close-ups); out-of-domain prompts drift towards anime characters; KID from 20 images is noisy.
License / use: Base model: CreativeML OpenRAIL-M (Stable Diffusion 1.5). Training images: Naruto frames from narutopedia.com via lambda/naruto-blip-captions (no licence stated on the dataset card; Naruto characters are copyrighted). For non-commercial research and teaching only.

Load: pipe.load_lora_weights("YashShevkar30/data266-lab2-naruto-lora_r8").

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for YashShevkar30/data266-lab2-naruto-lora_r8

Base model

Adapter

(754)

this model