Scholarlink

This study introduces a resource-efficient method for generating artwork in the traditional Madhubani style using Stable Diffusion XL (SDXL). Using the low rank adaptation (LoRA) technique, the model is fine-tuned with culturally relevant prompts and stylistic guidance to emulate the distinct characteristics of this Indian folk art form. To improve inclusivity and accessibility, prompt multilingual support is incorporated, covering Hindi, Bengali, Telugu, and English, through an automated translation mechanism, which retains the \(\texttt {&lt;madhubani-style&gt;}\) keyword to ensure consistency between languages. The training process is optimized for standard consumer-grade GPUs utilizing FP16 precision, CPU memory offloading, and a fixed Variational Autoencoder (VAE), enabling stable 1024 \(\times \) 1024 image generation. For evaluation, a Contrastive Language Image Pretraining (CLIP) based scoring method is employed to assess the semantic alignment between prompts and generated images. The findings indicate that the style remains remarkably consistent across languages, demonstrating that even minimal training can enable AI to effectively capture and preserve traditional art forms within digital media.</madhubani-style>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Parameter-Efficient Folk Art Generation: Fine-Tuning SDXL with LoRA for Madhubani Art Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Devashree Kute,
  • Anjali Naik

摘要

This study introduces a resource-efficient method for generating artwork in the traditional Madhubani style using Stable Diffusion XL (SDXL). Using the low rank adaptation (LoRA) technique, the model is fine-tuned with culturally relevant prompts and stylistic guidance to emulate the distinct characteristics of this Indian folk art form. To improve inclusivity and accessibility, prompt multilingual support is incorporated, covering Hindi, Bengali, Telugu, and English, through an automated translation mechanism, which retains the \(\texttt {<madhubani-style>}\) keyword to ensure consistency between languages. The training process is optimized for standard consumer-grade GPUs utilizing FP16 precision, CPU memory offloading, and a fixed Variational Autoencoder (VAE), enabling stable 1024 \(\times \) 1024 image generation. For evaluation, a Contrastive Language Image Pretraining (CLIP) based scoring method is employed to assess the semantic alignment between prompts and generated images. The findings indicate that the style remains remarkably consistent across languages, demonstrating that even minimal training can enable AI to effectively capture and preserve traditional art forms within digital media.