Japanese-specialized Stable Diffusion released

rinna Co., Ltd. has developed Japanese Stable Diffusion, which fine-tunes Stable Diffusion on images with Japanese captions.

The existing Stable Diffusion was trained based on an English dataset (LAION-2B-en), which meant that using Japanese prompts required a translation process, and it was difficult to accurately reflect Japan-specific cultural context or slang. For example, there was difficulty in visualizing Japan-specific concepts such as 'salary man.'

Japanese Stable Diffusion directly accepts Japanese prompts and generates images by understanding the culture, identity, and unique expressions of the Japanese-speaking world. The model has been released via Hugging Face and GitHub, and was implemented based on the Diffusers library.