kumuji

/

Sa2VA-i-4B

like 0

This PR significantly enhances the model card by:

• Corrected pipeline_tag: Changed from image-text-to-text to image-segmentation to accurately reflect the model's primary function of language-guided dense grounding and segmentation in images and video. This improves discoverability for users.
• Added descriptive tags: Included dense-grounding and referring-expression-segmentation for more precise categorization based on the model's core tasks.
• Enriched Content: The model card content has been substantially expanded by incorporating detailed information from the GitHub repository. This includes:
  • Explicit links to the paper (Sa2VA-i: Improving Sa2VA Results with Consistent Training and Inference) and the GitHub repository.
  • A visual teaser image.
  • Comprehensive sections on the model's overview, performance highlights, competition results, model zoo, quick start guide, and key technical improvements.
  • A consolidated "Citation" section for both Sa2VA-i and the original Sa2VA.
• Removed Redundant/Irrelevant Sections: The "File information" and the less comprehensive "Acknowledgement" sections have been removed to streamline the model card and adhere to best practices.

These changes provide a more complete, accurate, and user-friendly model card.