openclip

Pocket-Sized Multimodal AI for content understanding and generation across multilingual texts, images, and 🔜 video, up to 5x faster than OpenAI CLIP and LLaVA 🖼️ & 🖋️

Ultralytics fork of Apple MobileCLIP for fast image-text inference, training, evaluation, and an iOS demo.

Using Segment-Anything and CLIP to generate pixel-aligned semantic features.

Model merging, task-vector rebasin, and fine-tuning for vision and LLM models.

Clipora is a powerful toolkit for fine-tuning OpenCLIP models using Low Rank Adapters (LoRA).

I switched phones and found thousands of WhatsApp photos waiting for me, mostly memes and screenshots. This finds the ones worth keeping: search your pictures by describing them, and bin the rest into a quarantine you can undo.

Text-to-image search with OpenCLIP, Docker, Flask, Faiss, etc. and a basic front-end.

A doctor-assistive AI system that interprets medical knowledge and patient images simultaneously. It utilizes a Dual-Encoder architecture to cross-reference textbook theory with visual pathology, generating clinically grounded diagnoses.

An AI-powered computer vision system that automatically selects the best wedding photos from thousands of images.

Dynamic cluster-based data sampling for efficient and long-tail-aware vision-language model pre-training.

VALORA AI is a Multimodal Pricing Prediction Model that uses textual and visual data to make precise predictions on product prices

Masked Multi-Component Gated Decomposition Architecture

Local-first semantic image explorer powered by OpenCLIP embeddings and Qdrant vector search, enabling natural-language retrieval across screenshots and photos.

This work represents my Master's Degree Thesis in Applied Computer Science (Machine Learning and Big Data)

CLIPF: Contrastive Language-Image Pre-training with Word Frequency Masking — a frequency-based text masking strategy for efficient VLM pre-training (WACV 2026)

Feed-forward video-to-3D scene reconstruction using VGGT (CVPR 2025) with open-vocabulary semantic labeling via SAM 2.1 + CLIP. One command, no COLMAP, Apple Silicon + CUDA.

A redesigned CLIP architecture replacing the ViT encoder with modern CNN backbones (ConvNeXt V2) to improve efficiency and inference speed while maintaining strong vision-language alignment, enabling real-time use in robotics and edge applications.

An Edge-AI security pipeline for real-time violence detection and automated forensic reporting using YOLO, CoCa, and Qwen2.5-VL.

基于OpenCLIP的图像检索系统，支持文搜图、图搜图，结合FAISS能够通过用户上传图片库快速建立索引来实现图像检索