Japanese Stable CLIP ViT-L/16
Japanese Stable CLIP maps images and Japanese text into a shared embedding space. It supports image retrieval from text and similarity-based classification, rather than generating a response.
Model design and access
Training uses a SigLIP loss with translated CC12M and Japanese-captioned COCO data. The model can serve as a component in a generative system, but its own output is an embedding or similarity score. The card provides text and image encoding examples.Japanese Stable CLIP ViT-L/16 official model card
Practical use
Practical guidance. This model fits Japanese image search where query and image representations need to be compared. Evaluate retrieval ranking on the actual image collection. Similarity is relative to the candidates and preprocessing, so do not interpret a high score as proof that every detail in a caption is present.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Japanese Stable CLIP |
| Image encoder | ViT-L/16, initialized from an AugReg model.Japanese Stable CLIP ViT-L/16 official model card |
| Text encoder | 12-layer BERT with a Japanese RoBERTa tokenizer.Japanese Stable CLIP ViT-L/16 official model card |
| License | Stability AI Community License.Japanese Stable CLIP ViT-L/16 official model card |
Related models
- Japanese Stable Diffusion XL
Stability AI · Stable Diffusion
Japanese Stable Diffusion XL adapts SDXL 1.0 for Japanese prompts and cultural concepts.
Launch date not confirmedDownloadable - Japanese Stable VLM
Stability AI · Japanese Stable VLM
Japanese Stable VLM generates Japanese descriptions and answers from an image, with optional text instructions.
Launch date not confirmedDownloadable