Int8 dynamic-quantized ONNX export of Falconsai/nsfw_image_detection (ViT-base 224, Apache-2.0), mirrored so an upstream change cannot silently alter the Generate AI Images extension's safety check.

Contract: input pixel_values [B,3,224,224] float32 CHW — bilinear resize to 224², /255, then (x−0.5)/0.5; output logits [B,2] → softmax; id2label {0: normal, 1: nsfw}.

Quantization validated against the fp32 torch model: scores agree to 3–4 decimals on a benign photorealistic test set. Replaces the previous AdamCodd-based mirror, which scored fully-clothed photorealistic portraits at 0.6–0.85 nsfw (false-positive band overlapping the block threshold); this model scores the same set at ≤0.005.