Stable Diffusion on Mac

[Stable Diffusion] Generating Popular Characters with etherPDXL

#StableDiffusion

What is etherPDXL?

etherPDXL is a high-quality image generation model based on SDXL. It excels at creating realistic and anime-style images, and by applying LoRA, it can faithfully reproduce specific characters. This time, we applied four different LoRAs and actually generated characters.

What is LoRA (Low-Rank Adaptation)?

LoRA (Low-Rank Adaptation) is a technique that applies pre-trained characters or styles to Stable Diffusion. It allows you to add specific characters or art styles to existing trained models (such as Stable Diffusion 1.5 and SDXL), making it easy to reproduce your favorite anime characters or original styles.

Features of LoRA

  • Lightweight and fast → Because it only adds specific features instead of training the entire model, it loads quickly and has low memory usage.
  • Add-on to existing models → Easily reproduce your favorite anime characters or art styles simply by applying a LoRA with the desired characteristics.
  • Adjustable LoRA intensity → Adjustable from 0.5 to 1.2, allowing you to freely change the degree of application.

How to Apply LoRA

In Stable Diffusion WebUI (AUTOMATIC1111), you can apply LoRA as follows:

  1. Place the LoRA file (.safetensors) in the following folder:

    stable-diffusion-webui/models/Lora/

  2. Restart the WebUI and select the LoRA in the "Extra Networks" tab.

  3. Add <lora:LoRA name:strength> to the prompt.

    <lora:zerotwo_offset:0.9>

LoRA 1: Okarun (Okaruto-kun / Dandadan)

LoRA Name: Okarun Dandadan [IL]

This LoRA reproduces Okaruto-kun (Okarun), the protagonist of Dandadan.

Generated Image

Dandadan-Okarun

Observations

  • Because etherPDXL is SDXL-based, applying LoRA with an intensity of 1.0 or higher results in more stable application.

  • Facial details tend to become slightly too realistic in an SDXL style; therefore, adjust the Hires Fix Denoising Strength to 0.4–0.5.

LoRA 2: Aqua / Megumin (KonoSuba LoRA)

LoRA Name: konosuba_collection_v2

This LoRA can generate KonoSuba (KonoSuba: God's blessing on this wonderful world!) characters. Supports multiple characters, including Aqua, Megumin, and Darkness.

Generated Images

KonoSuba-Aqua

KonoSuba-Megumin

Observations

  • Because etherPDXL is SDXL-based, the application of SD1.5 LoRAs is weaker; therefore, an intensity of 0.8 or higher is appropriate.
  • Eye depiction tends to be in an older anime style, so specify (old anime style:1.3) in the Negative Prompt.
  • Adding (modern anime style:1.3) to the prompt improves the quality for a more modern anime look.

LoRA 3: Zero Two (Darling in the FranXX)

LoRA Name: zerotwo_offset

Reproduces Zero Two from Darling in the FranXX.

Generated Image

Zero Two

Observations

  • In etherPDXL, an intensity of 0.9–1.1 is optimal.
  • Because it's SDXL, the effect of the LoRA is weaker than with SD1.5; set Clip Skip to 2.
  • Zero Two's characteristics (pink hair + red uniform) are easily produced; however, an SF-style background makes her stand out even more.

LoRA 4: Frieren (Frieren: Beyond Journey’s End)

LoRA Name: sousou_no_frieren_collection_v1.2

This LoRA can generate Frieren, the protagonist of Frieren: Beyond Journey’s End.

Generated Image

sousou_no_frieren

Observations

  • With etherPDXL's SDXL base, around 0.85 is ideal for application.
  • Adding a "magic effect" background enhances the Frieren-like atmosphere.
  • Adjust the LoRA intensity to 0.85–1.0 to improve the accuracy of hair color and costume details.