Initializing...
Predictions
Ready to analyze
This demo shows where a vision model looks. Its main view is the Label-free Activation Map (LaFAM): a heatmap of the regions the model responds to most, computed without choosing a class, so it also works for models trained without labels (Karjauv & Albayrak, 2024; blog post).
The demo also reads the classifier region by region. A classifier divides a photo into a grid of regions, describes each region, and averages these descriptions into one prediction. Because averaging is the last step, its answer can be read for every region, and the average of these answers is exactly its prediction. This shows where it sees each class (CAM, class activation map) and lets it find several objects in one image (Karjauv, 2026).
Models with another pooling, such as CLIP RN50x4 (attention pooling), give their own prediction; their regions are each classified on their own, as in MaskCLIP (Zhou et al., 2022), so the regions no longer average to the prediction.
The pathology model, Hibou-B, expects H&E tissue at 0.5 µm per pixel (20×). It takes a 224 px square, so it sees 112 µm of tissue, and at other scales its maps are unreliable. The breast examples are 112 µm tiles at 0.25 µm per pixel (448 px), which the demo resizes to 224 px. An image you upload is cropped to its centre square and resized to 224 px whatever its pixel size, so upload a square tile about 112 µm wide. The tissue head of Hibou-B was trained with one label per tile on breast cancer tiles (BCSS) with no lobular case among them.
Everything runs in your browser, and your images never leave your device.
Examples: breast tissue from BCSS (TCGA-BRCA; CC0); apple photos from Goh et al., 2021 (CC-BY 4.0). Model licences: see the model files.
Pick an example, upload an image or video, or use your camera. Pause a video to interact with it.
If you encounter issues, use the hard reset button below to clear cache and manually reload the page.
Description goes here.