LoRA allows for changing styles through additional training without altering the base generative AI model, making it easy to implement. This time, I experimented with LoRA using the extension capabilities of Stable Diffusion. Since it&#39;s included as standard, it can be utilized by importing models from Hugging Face or Civitai. <a href="https://github.com/AUTOMATIC1111/stable-diffusion-webui" rel="noopener noreferrer" target="_blank">https://github.com/AUTOMATIC1111/stable-diffusion-webui</a> <a href="https://huggingface.co/" rel="noopener noreferrer" target="_blank">https://huggingface.co/</a> <a href="https://civitai.com/" rel="noopener noreferrer" target="_blank">https://civitai.com/</a> Choosing LoRA as the base for the prompt adds commands like `&lt;hoge&gt;` for LoRA, making it incredibly easy to use just by that addition. The LoRA I utilized this time was... <a href="https://civitai.com/models/116730?modelVersionId=126394" rel="noopener noreferrer" target="_blank">https://civitai.com/models/116730?modelVersionId=126394</a> base model is like, <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/withoutlora.png"/> The same text prompt with LoRA,<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/lora.png"/> There is hardly any difference in inference speed.<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/lora_benchmark.png"/>

Trying out LoRA with Stable Diffusion Web UI.

Yuichiro Minato