sd15-lora-album-covers-rock

SD 1.5 LoRA — Rock & Metal album-cover style (trigger: rocov)

LoRA adapter for stable-diffusion-v1-5/stable-diffusion-v1-5, fine-tuned on kramp/album-covers-rock-metal-blip: 6,000 album covers from eong/20k-Album-Covers-within-20-Genres filtered to the DeathMetal, DoomMetal, HeavyMetal, PsychedelicRock, Punk and Rock classes, captioned with BLIP (Salesforce/blip-image-captioning-large) and prefixed with the trigger phrase.

Usage

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights("kramp/sd15-lora-album-covers-rock")

image = pipe(
    "rocov, album cover art, a chrome skull on a black background",
    num_inference_steps=30, guidance_scale=7.0,
).images[0]
image.save("cover.png")

Prefix prompts with rocov, album cover art, to activate the style; without the rocov trigger the pipeline behaves like base SD 1.5.

Training details

• Script: diffusers examples/text_to_image/train_text_to_image_lora.py (v0.40.0, run verbatim)
• LoRA rank 16 (alpha = 16), target modules to_k/to_q/to_v/to_out.0 on the UNet
• lr 1e-4 (constant), resolution 512, fp16, gradient checkpointing, seed 42
• batch 8 × gradient accumulation 2, 2,000 optimizer steps ≈ 5.3 epochs
• Final training loss ≈ 0.18; loss curve on the training dashboard
• Source covers are 300×300 upsampled to 512 — expect some softness at output edges

Validation images (generated during training)

Held-out samples

Limitations & bias

• Trained only on covers from the 6 rock/metal classes — outputs drift toward dark metal/punk aesthetics even for neutral prompts.
• BLIP captions are generic scene descriptions; any text/logo on real covers is not learned and will not be legible at 512px.
• SD 1.5's safety checker occasionally flags dark metal-style outputs and returns a black image — if that happens, retry with a different seed.