DINOv3 ViT-L · Live Inference
Real forward + backward passes on CPU: predictions, Grad-CAM, and robustness probes.
analytics97.2% orig · 94.9% cleanWeapons · 9-class300M (LoRA)In this workspace:Attention mapsFGSM attackCorruptionsOOD signal
Recognizes9 weapon categories9 classes · bazooka, grenade launcher, rifle, shotgun, smg, sniper, knife, sword, handgun
Accuracy97.2% orig · 94.9% clean97.2% on the original split, 94.9% on the clean (leakage-proof) split
Works best onA photo of one of the nine weapon typesResized to 224×224 color, ImageNet-normalized.
warning
Expect confident mistakes outside its scope
This model only knows nine weapon classes and has no "no weapon" option. Shown an ordinary object — an animal, a vehicle, a face — it will still return a weapon label, sometimes confidently. We measured this directly: 27 of 27 non-weapons were labelled as weapons. Treat the confidence value, not just the label, as the signal.
Top-5 Predictions
Upload an image and run analysis to see predictions.
Input
Activation Map
LOW→HIGH
Layer Source
—
Focus Metric
—
Latency
—
Confidence (top-1)
—
Run analysis to compute
Inference Latency
—
CPU forward pass