HZ Wallpaper — Minimal Abstract 3D SDXL LoRA

A style LoRA trained to create premium minimalist abstract wallpapers with smooth 3D forms, flowing curves, clean geometric structures, soft gradients, subtle depth, and elegant lighting.

The LoRA is designed to work especially well for desktop wallpapers, abstract backgrounds, modern UI backgrounds, and clean premium visual designs.

It works best when the prompt is kept relatively simple. The LoRA already contributes much of the visual style, so you do not need a huge prompt to get good results.

Base Model

Juggernaut Ragnarok XL / SDXL

Trigger Words

hz_wallpaper

Positive Prompt:

hz_wallpaper, abstract, minimalist abstract 3D wallpaper, smooth flowing forms, soft gradients, clean composition, subtle depth, elegant lighting

Negative Prompt:

text, logo, watermark, signature, people, face, human, objects, clutter, blurry, grainy, low resolution, noise, oversaturated, muddy colors, distorted geometry, bad composition

⭐ Recommended Settings

These are the settings I recommend starting with:

LoRA Strength: 0.3 – 0.7

Steps: 25–30

CFG: 5.5–7

Sampler: dpmpp_2m_sde

Scheduler: Karras

Resolution: 1024×1024

All Resolutions:

• 1:1 – 1024 x 1024 (Native base)

Portrait & Vertical

• 4:5 – 896 x 1152 (Ideal for character portraits)

• 2:3 – 832 x 1216 (Standard photo portrait)

• 9:16 – 768 x 1344 (Perfect for mobile wallpapers and stories)

• 1:2 – 640 x 1536 (Cinematic vertical)

Landscape & Widescreen

• 5:4 – 1152 x 896 (Slightly wide)

• 3:2 – 1216 x 832 (Classic landscape photography)

• 16:9 – 1344 x 768 (Standard widescreen / desktop wallpaper)

• 2.4:1 – 1536 x 640 (Cinematic panorama)

Denoise: 1.0

⚠️ Important Notes

The LoRA does not continue learning during generation.

Once the LoRA has been trained, its weights are fixed. Changing the LoRA strength only changes how strongly those learned weights influence the base model.

The recommended strength range of 0.3–0.7 is therefore an inference recommendation, not a training setting.

Also, results can vary depending on the base model, sampler, seed, resolution, and prompt. The settings above are the recommended starting point, not hard requirements.