CLIP-embeddings — семантический поиск изображений

★ 6.8 · general

clip-aware-embeddings is a Claude Code skill that handles semantic image-text matching with CLIP and automatically routes tasks to better alternatives when CLIP falls short. It activates on queries about image similarity, semantic search, zero-shot classification, and embeddings, then applies a built-in decision tree: CLIP for broad category search and similarity matching, DETR or Faster R-CNN for counting objects, specialized models for fine-grained identification like celebrities or car models, GQA/SWIG for spatial reasoning, and PC-CLIP or DCSMs for compositional queries such as attribute binding. The skill requires transformers, pillow, torch, and sentence-transformers, and can leverage Firecrawl and Hugging Face MCP integrations for research and model documentation. It targets developers and ML engineers who need working code patterns for image-text tasks without the common pitfall of applying CLIP to problems its architecture cannot solve.