Model Card for Glyphress/Qwen_Image_Korean_Font_LoRA 🔥

This model card documents the Korean Font LoRA developed by Glyphress for Qwen/Qwen-Image, focused on high-quality Hangul font generation, editing, and stylization using LoRA fine-tuning.

Korean font generation and editing with LoRA fine-tuning.

---

Model Details

Model Description

This repository provides a LoRA (Low-Rank Adaptation) module fine-tuned on top of Qwen/Qwen-Image, specialized for Korean typography and glyph-style generation.

The primary goals are:

• Improve legibility and consistency of Korean text generated by Qwen-Image
• Enable logo-like, stylized rendering of Korean titles and phrases
• Provide a creative tool for font and branding exploration in Korean contexts

This LoRA merges:

• A font-focused LoRA, trained for structural clarity and legibility
• A style-focused LoRA, trained for visual mood and stylistic variation

The model does not include the base weights.
It must be applied on top of Qwen/Qwen-Image.

• Developed by: Glyphress
• Model type: LoRA adapter for diffusion (image generation)
• Languages: Optimized for Korean, but supports English prompts for style control
• License: openrail++
• Finetuned from: Qwen/Qwen-Image

Model Sources

• Model Repository:
  Glyphress/Qwen_Image_Korean_Font_LoRA
• Hugging Face Space:
  Qwen Image Korean Font LoRA 🔥
• Author / Profile:
  https://huggingface.co/Glyphress
• GitHub (Author):
  https://github.com/Sumin0510

---

Uses

Intended / Direct Use

This LoRA is designed for:

• Korean logo & typography generation
• Font style exploration (modern, minimal, logo-like)
• Poster / title / banner creation
• Creative content prototyping

Typical users may include:

• Designers
• Artists
• Content creators
• Researchers exploring generative typography

Downstream Use

May be integrated into:

• Web-based UI / Spaces for Korean font exploration
• Typography ideation tools
• Automated design systems generating title graphics

Out-of-Scope Use

Not intended for:

• Fraud, trademark impersonation, counterfeiting
• Generating official documents, certificates, IDs
• Disallowed uses under openRAIL++ or Qwen policies
• Safety-critical applications

May perform poorly on:

• Very long sentences
• Dense multi-line layouts
• Complex backgrounds
• Non-Korean scripts

---

Bias, Risks, and Limitations

Known Limits

• Legibility decreases when:
  • Background is complex
  • Font size is small
  • Prompts are long
• May produce inaccurate glyph strokes
• Style bias toward:
  • Logo-like, modern styles
  • High contrast compositions

Intellectual Property Note

Training data used commercially-free fonts only
and did not include real corporate logos or trademarks.

Nevertheless, users must ensure compliance with:

• Copyright
• Trademark laws
• Font license obligations

Recommendations

For production use, users should:

• Manually review outputs
• Avoid close imitation of existing trademarks
• Avoid using generated logos without legal review

---

How to Get Started with the Model

Basic Usage (Diffusers)

from diffusers import DiffusionPipeline
import torch

base_model = "Qwen/Qwen-Image"
lora_repo = "Glyphress/Qwen_Image_Korean_Font_LoRA"
weight_name = "Glyphress.safetensors"

pipe = DiffusionPipeline.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    use_safetensors=True,
    trust_remote_code=True,
    device_map="balanced",
)

pipe.load_lora_weights(
    lora_repo,
    weight_name=weight_name,
)

pipe.to("cuda")

prompt = "Korean text '한국대학교', black font color, white background, simple modern logo style"

image = pipe(
    prompt=prompt,
    num_inference_steps=20,
    guidance_scale=3.0,
    height=512,
    width=512,
).images[0]

image.save("example.png")