Sakana unveils Japanese image model

Sakana AI just released EvoSDXL-JP, a speedy text-to-image generator specifically tailored for Japanese-style images, applying the startup’s ‘evolutionary model merging’ technique beyond language models.

 

The details:
  • EvoSDXL-JP was built by merging a Japanese image model with an English one, then merging the result with a faster diffusion model for a 10x speed increase.
  • Sakana’s model merging involves ‘breeding’ hundreds of model generations, with the most successful ones becoming ‘parents’ for the next.
  • The model supports Japanese text prompts and can generate culturally relevant images, aiming to make genAI more accessible in the country.
Why it matters: Sakana’s evolutionary model merging technique was already demonstrating unique ways to improve LLMs — and EvoSDXL-JP now shows the same process can apply across boosting image generation.

Link website: https://sakana.ai/evosdxl-jp/

Leave a Reply

Your email address will not be published. Required fields are marked *