Feature visualization is part of my daily routine as a computer vision researcher. But doing it across different model providers was always a mess. So this weekend, I finally fixed it 🧩 Every model lives somewhere different, torch.hub, timm, HuggingFace, some random GitHub repo, and each one needs its own glue code just to pull out and visualize feature maps. I got tired of rewriting that glue every single time. 🤖 So I built FeatLens: a model-agnostic framework to see what any vision model actually encodes. 🔹 Load from anywhere — timm, HuggingFace, torch.hub (like V-JEPA), an external repo, or your own nn.Module (yes, CNNs too) 🔹 Any layer, laid out as a clean model × layer grid 🔹 Color the features by PCA, cosine similarity, k-means, foreground, saliency, or attention-rollout 🔹 Match patches across two images, batch a whole folder, or sweep a video clip Just swap the models, and watch the features change completely. It's oddly satisfying to look at. This was a weekend project I built with my friend Claude, from core design to docs, tests, and CI. It's open source, and you can just pip install featlens ⭐ GitHub: https://lnkd.in/dGvb7uNi 📖 Docs: https://lnkd.in/dQMV9-Bc 🤗 Live demo: https://lnkd.in/d6HeRJk6 If it's useful in your work, it's citable, and contributions are very welcome. Which model would you want to peek inside first? 👀

FeatLens

Visualize feature maps from any vision model effortlessly.

Visualize feature maps from any vision model effortlessly.