|
| 基于改进稳定扩散模型的蓝印花布纹样生成 |
| Blue calico pattern generation based on improved stable diffusion model |
| 投稿时间：2024-03-27 修订日期：2024-04-18 |
| DOI： |
| 中文关键词: 蓝印花布 稳定扩散模型 低秩自适应算法 判别网络 |
| 英文关键词:blue calico Stable diffusion Low rank adaptive algorithm Discriminative network |
| 基金项目:浙江省自然科学基金资助项目 |
|
| 摘要点击次数: 41 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对我国非物质文化遗产蓝印花布纹样的数字化传承与创新技术缺乏问题，提出一种基于改进的稳定扩散模型的蓝印花布纹样生成方法，通过文生图和图生图技术实现蓝印花布单纹样和多纹样的主动生成。对于蓝印花布纹样数据集较少，无法训练出大模型难题，提出将稳定扩散模型和低秩自适应算法（LoRA）微调网络结合，训练出具有蓝印花布特征的LoRA微调网络；对于稳定扩散模型输出具有随机性问题，提出将稳定扩散模型与判别网络相结合，对生成的图片进行判断，筛选出符合蓝印花布特征的纹理图片。实验结果表明，通过关键提示词或图片，可以生成具有纹样语义信息和图片特征的蓝印花布新纹样。 |
| 英文摘要: |
| Aiming at the problem of the lack of digital inheritance and innovation technology of blue calico patterns of China’s intangible cultural heritage,this paper proposed a blue calico pattern generation method based on the improved stable diffusion model,which realized the active generation of single and multiple blue calico patterns through text-generated graphics and graph-generated graphics. For the problem that there are few blue calico pattern datasets,and it is difficult to train a large model,this paper proposed to combine the stable diffusion model and Low-Rank Adaptive Algorithm(LoRA)fine-tuning network to train the LoRA fine-tuning network with blue calico patterns. For the problem that the output of the stable diffusion model is random,this paper proposed to combine the stable diffusion model and the discriminant network to judge the generated pictures and select the texture pictures that conform to the characteristics of blue calico patterns. Experimental results show that new blue calico patterns with semantic information and image features can be generated through key prompt words or pictures. |
| 下载PDF阅读器 |
| 关闭 |