plesk8er/EmoNet-Face-ONNX

EmoNet-Face ONNX for TouchDesigner

Single-graph ONNX conversion of laion/Empathic-Insight-Face-Large (40 emotion regression heads from the EmoNet-Face paper) on top of the google/siglip2-so400m-patch16-384 image encoder, packaged for the Convivial Studio EmoNet-Face component for TouchDesigner. Weights are unchanged; the 40 MLP heads are stacked into one pass and the SigLIP2 normalization is baked in.

File	What it is
emonet_face[_fp16].onnx	image float32 [1,3,384,384] RGB 0..1 face crop in; scores [1,40] (EmoNet-Face scale, about 0..7) and embedding [1,1152] out
emonet_face_meta.json	state names, channel keys, per-state neutral means from the original repo
face_detection_full_range.onnx	MediaPipe BlazeFace full-range face detector, converted from TFLite

Licenses and credit

• Empathic-Insight-Face heads: CC-BY-4.0, LAION. Schuhmann et al., EmoNet-Face: An Expert-Annotated Benchmark for Synthetic Emotion Recognition, NeurIPS 2025, arXiv:2505.20033.
• SigLIP2 image encoder: Apache-2.0, Google.
• BlazeFace face detector: Apache-2.0, Google MediaPipe.
• Conversion and packaging: Convivial Studio.

Intended use

Creative and research use. The model was trained on synthetic portraits with clear expressions; treat its scores as soft signals, ideally relative to a personal neutral baseline. As stated by the EmoNet-Face authors, emotion recognition in workplaces and educational institutions is excluded (EU AI Act Art. 5(1)(f)).