user@devops:~$ cat README.md
CLIP Multimodal
# Description
Multimodal project with OpenAI CLIP (Contrastive Language-Image Pre-training) that bridges images and text in the same 512-dimensional vector space. The pipeline loads 20 real images from picsum.photos, encodes them with CLIP's Vision Transformer (ViT-B/32), and performs zero-shot classification across 74 text-defined categories. Also includes semantic image search by text (10 queries over 20 images with top-5 results), a text-image similarity matrix, and t-SNE visualization of the multimodal embedding space. CLIP classifies images without ever having seen them during training, simply by comparing image embeddings with textual descriptions.
# Key features
$ Zero-shot classification across 74 categories with OpenAI CLIP (ViT-B/32)
$ Semantic image search via natural language queries with cosine similarity
$ 512-dimensional multimodal embeddings: images and text in the same vector space
$ Text-image similarity matrix with heatmap for 10 semantic queries × 20 images
$ t-SNE projection of the multimodal space showing images and textual references
$ 7 visualizations: t-SNE, heatmap, semantic search grid, confidence, distribution, examples
$ Real dataset: 20 diverse images from picsum.photos (nature, objects, people, animals)
# Gallery
# Technologies used
