facebookresearch/convnext

ConvNeXt is a modern image recognition model that works like a traditional camera-based AI, but rebuilt from the ground up using ideas from recent breakthroughs. The core benefit is that it achieves top-tier accuracy on real-world image classification tasks while remaining simple, efficient, and practical to deploy. Instead of using trendy attention mechanisms (the approach that powers large language models), ConvNeXt sticks with straightforward convolutional layers, making it faster and more resource-friendly for common computer vision work. At a high level, the model takes an image as input and outputs what it thinks the image contains, like "this is a dog" or "this is a cat." It comes in different sizes (tiny, small, base, large, extra-large), so you can pick one that matches your hardware and speed requirements. The repository includes pre-trained versions of these models that have already been trained on millions of labeled images, so you don't need to train from scratch. You can either use them as-is to classify new images, or fine-tune them on your own image collection to teach them specialized tasks. Researchers and machine learning engineers would use this for tasks like building image search systems, automating photo sorting, detecting objects in videos, or powering visual inspection in manufacturing. A startup building a document scanning app, for example, could use ConvNeXt to automatically classify document types. A research lab studying medical imaging could fine-tune it to spot anomalies in X-rays. The repository provides everything needed: the model code, training recipes, evaluation tools, and ready-to-download pre-trained weights. The project notably chooses simplicity over complexity. While other modern models chase cutting-edge techniques, ConvNeXt proves that a well-designed classical approach, combining traditional convolutions with modern training tricks, can match or beat fancier alternatives. The README includes performance numbers showing accuracy across different model sizes and input resolutions, along with downloadable trained models so you can start using it immediately without expensive training runs.