A LoRA fine-tune of SDXL 1.0 trained on black metal album covers, band logos, and zine art. Generates high-contrast dark art in the visual language of black metal: fog-drenched forests, burning churches, corpse paint portraits, illegible spiky logos, frozen mountain landscapes, and occult imagery — the raw, monochrome aesthetic from the early Norwegian scene through modern atmospheric and DSBM.

Trigger word: black-metal-art

Example Generations

Landscapes & Scenes

Dark Forest (0.7)	Burning Church (0.7)	Frozen Mountain (0.7)
Forest	Church	Mountain

Graveyard (0.7)	Castle Ruins (0.6)	Frozen Lake (0.6)
Graveyard	Ruins	Lake

Mist & Branches (0.7)	Northern Lights (0.6)	Burning Church (0.6)
Mist	Aurora	Church 0.6

Portraits & Figures

Corpse Paint (0.7)	Corpse Paint (0.5)	Hooded Figure (0.7)	Warrior (0.6)
Corpse 0.7	Corpse 0.5	Hooded	Warrior

Logos & Typography

Thorny Logo (0.8)	Thorny Logo (0.6)	Fire Logo (0.7)
Logo 0.8	Logo 0.6	Logo Fire

Black metal logos are intentionally illegible — garbled, spiky text output is a feature, not a bug. Higher LoRA scales (0.7–0.8) produce more authentic logo aesthetics.

Album Cover Compositions

Wolf & Moon (0.7)	Horned Skull (0.7)	Raven & Skull (0.6)	Goat (0.7)
Wolf	Skull	Raven	Goat

LoRA Scale Comparison

The same prompt ("dark forest, fog, dead trees") at different LoRA scales:

Scale 0.5	Scale 0.7
Forest 0.5	Forest 0.7

Training Progression

The same prompt at different training checkpoints (forest scene, LoRA scale 0.7):

Step 2250	Step 2500	Step 2750
2250	2500	2750

Step 3000	Step 3250	Step 3500 (Final)
3000	3250	3500

Usage

With Diffusers

python

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
)
pipe.load_lora_weights("cahlen/black-metal-art-sdxl-lora")
pipe.set_adapters(["default_0"], adapter_weights=[0.7])
pipe.to("cuda")

image = pipe(
    prompt="black-metal-art, dark forest, fog, dead trees, high contrast, monochrome",
    negative_prompt="blurry, photorealistic, photo, smooth gradients, 3d render, watermark, clean, modern, digital",
    width=1024,
    height=1024,
    num_inference_steps=30,
    guidance_scale=7.5,
).images[0]

image.save("black_metal_art.png")

With ComfyUI / A1111

Load pytorch_lora_weights.safetensors as a standard SDXL LoRA. Use trigger word black-metal-art in your prompt.

Recommended Settings

Parameter	Recommended	Notes
LoRA scale	0.5–0.8	Higher scales = stronger style. See guide below.
Resolution	1024x1024 or 768x1024	Square for album covers, portrait for scenes
CFG / Guidance	7.5	Standard SDXL guidance
Steps	25–30	Diminishing returns past 30
Sampler	Euler a, DPM++ 2M Karras	Standard SDXL samplers

LoRA Scale Guide

Scale	Effect	Best For
0.5	Moderate style influence	Subtle dark atmosphere, blending with other styles
0.6	Strong style, good detail	Landscapes, portraits, album cover compositions
0.7	Full black metal aesthetic	Forests, churches, corpse paint — the default sweet spot
0.8	Very strong, maximum style	Logos, abstract dark imagery, maximum raw aesthetic

Text & Logo Generation

Unlike most LoRAs, garbled text is welcome with this model. Black metal band logos are famously illegible — the spiky, branch-like letterforms that SDXL generates naturally match the genre aesthetic. For logo generation, use higher LoRA scales (0.7–0.8) and prompts like:

black-metal-art, illegible spiky band logo, branches and thorns
black-metal-art, band logo, flames, inverted cross, black background

Prompt Tips

Always start with the trigger word: black-metal-art
Landscape prompts work great: forests, mountains, frozen lakes, graveyards, ruins
Monochrome keywords reinforce the style: high contrast, monochrome, black and white
Atmospheric keywords: fog, mist, bleak, desolation, frozen, winter
Best subjects: dark forests, burning churches, corpse paint, mountains, wolves, ravens, skulls, occult symbols, band logos
Album cover format: Use 1024x1024 square resolution

Training Details

Parameter	Value
Base model	stabilityai/stable-diffusion-xl-base-1.0
Framework	SimpleTuner
LoRA rank / alpha	32 / 32
Optimizer	Prodigy (auto-adjusting LR, initial 1.0)
Batch size x Accum	1 x 4 (effective batch 4)
Max steps	3,500
Checkpoints	Every 250 steps (14 total)
Precision	bf16
SNR gamma	5
Noise offset	0.05
Caption dropout	0.20

Dataset

Metric	Value
Sources	MusicBrainz / Cover Art Archive, Internet Archive
Raw scraped	7,618 black metal album covers, zine scans, and flyer art
Curated	Top 200 by programmatic scoring (entropy, contrast, B&W fraction, edge density, aspect ratio, resolution)
After preprocessing	400 images (200 original + 200 B&W augmentation)
Repeats	2
Effective samples/epoch	800
Total epochs	~17.5
Caption strategy	Instance prompt only (`black-metal-art`) — style LoRAs learn better when captions describe content only

Why SDXL over FLUX

SDXL 1.0 (3.5B parameters) was chosen over FLUX (12B+) because FLUX's strong photorealistic prior overwhelms rank-32 style LoRAs. SDXL's smaller parameter count is more amenable to specialized art styles at practical LoRA ranks.

Style Reference

This LoRA captures the visual language of black metal art:

High-contrast monochrome — dark, raw, uncompromising
Norwegian black metal scene — the iconic aesthetic of Mayhem, Burzum, Darkthrone, Emperor
Frozen landscapes — snow-covered forests, mountains, desolate winter scenes
Burning churches — the infamous visual motif of early Norwegian black metal
Corpse paint — the theatrical black and white face paint
Illegible logos — spiky, branch-like letterforms, intentionally unreadable
Occult imagery — pentagrams, inverted crosses, goat heads, ritual scenes
Album cover compositions — wolves, ravens, skulls, ancient ruins

License

This LoRA inherits the CreativeML OpenRAIL-M license from the SDXL base model.

Acknowledgments

Training data sourced from public digital archives (MusicBrainz / Cover Art Archive, Internet Archive). This LoRA is a tribute to the visual art tradition of black metal and the artists who created this iconic aesthetic.