CognitiveOS/vision

face-api weights (gguf + safetensors)

Pre-trained model weights from vladmandic/face-api at commit 189226d63aabb48cb40776fd1c453ebc0fa722f1, converted losslessly to portable safetensors and gguf (float32) formats by the CognitiveOS-Labs/tfjs-weights-to-gguf reproducible pipeline.

Models

File	Tensors	Purpose
tiny_face_detector_model.gguf	19	Tiny face detection
face_landmark_68_model.gguf	49	68-point landmarks
face_landmark_68_tiny_model.gguf	28	Tiny 68-point landmarks
face_recognition_model.gguf	117	Face embedding (128-d)
ssd_mobilenetv1_model.gguf	151	SSD face detection
age_gender_model.gguf	51	Age / gender
face_expression_model.gguf	49	Facial expression

Every model is also provided as .safetensors and both formats carry a *-mapping.json recording each tensor's name, shape, dtype, data location, and the original TF.js quantization block (source traceability).

Format notes

• Weights only. The face-api repo has no model.json — the network topologies live in its TypeScript sources under src/. These artifacts preserve the original TF.js tensor names verbatim.
• All tensors are float32 (the dequantized values produced by the TF.js loader), except 15 scalar int32 constants in ssd_mobilenetv1_model (postprocessor shape parameters).
• The gguf files use generic packing (metadata keys general.architecture, general.name, general.source, general.source.commit, general.alignment, general.file_type; GGML dim order; GGML_TYPE_F32=0, GGML_TYPE_I32=26). They are structurally valid GGUF v3 but use a custom general.architecture, so llama.cpp will not load them yet — a runtime for these architectures is the documented follow-up (see the face-recognition CGP repo).
• Reference loaders and the full bitwise-validation suite live in the CognitiveOS-Labs/tfjs-weights-to-gguf repo under scripts/.

Consumed by

• CognitiveOS-Labs/face-recognition — CognitiveOS patch declaring face_recognition_model.gguf as its remote wide-model weight (downloaded at cpm install time).
• CognitiveOS-Labs/face-recognition-tune — cpm tune LoRA trainer over the frozen base weights.

Source & license

Weights derived from vladmandic/face-api (MIT), commit 189226d63aabb48cb40776fd1c453ebc0fa722f1, provided "as-is". Conversion tooling in the tfjs-weights-to-gguf repo is MIT.

The recognition backbone was trained on VGGFace2-like data by the face-api project; generic face recognition works out of the box, while recognizing specific people is an enrollment (embedding-gallery) operation.