@@ -8,9 +8,9 @@
 
 face-detect.cpp is a from-scratch C++17 port of the [insightface](https://github.com/deepinsight/insightface) buffalo face-recognition pipeline, built on [ggml](https://github.com/ggml-org/ggml). It runs the full chain (SCRFD detection, 5-landmark similarity-transform alignment to 112x112, ArcFace embedding) from a single self-contained GGUF file, with no Python, no PyTorch, and no onnxruntime at inference time. The honest headline is not raw CPU speed (on CPU it trails onnxruntime's MLAS conv kernels, and we say so below): it is simplicity, breadth, and bit-exact parity. One small `libfacedetect.so` does detect, align, recognize, age/gender, anti-spoof, and dense landmarks across six model packs, with detector boxes and landmarks matching insightface to within 1 px and the recognition embedding matching to cosine 1.000000. It is the native replacement for LocalAI's Python `insightface` backend, consumed through a flat C ABI (`include/facedetect_capi.h`) via purego / dlopen.
 
-![face-detect.cpp: live face recognition across 10 real faces, detect plus 5 landmarks plus recognize plus age and gender, one binary, no Python](benchmarks/media/face_carousel.gif)
+![face-detect.cpp: live face recognition across 10 famous paintings, detect plus 5 landmarks plus recognize plus age and gender, one binary, no Python](benchmarks/media/face_carousel.gif)
 
-Live face recognition across 10 real faces: detect, 5-landmark align, recognize, and age and gender, all from one self-contained binary with no Python ([full clip](benchmarks/media/face_carousel.mp4), [square version for social](benchmarks/media/face_carousel_square.mp4)). The full per-face output on a single photo:
+Live face recognition on 10 famous public-domain paintings (the Mona Lisa, Vermeer, Rembrandt, Van Gogh and more): detect, 5-landmark align, recognize, and age and gender, all from one self-contained binary with no Python ([full clip](benchmarks/media/face_carousel.mp4), [square version for social](benchmarks/media/face_carousel_square.mp4)). The age and gender are the model's literal reading of a painting, which makes for some charming honest mis-reads. The full per-face output, here on the Mona Lisa:
 
 <p align="center">
   <img src="benchmarks/media/face_demo.png" width="92%" alt="the full pipeline output on one photo: detect box, 5 landmarks, recognition, age and gender, and verify">