---
language:
- es
- en
license: cc-by-nc-nd-4.0
pipeline_tag: image-classification
tags:
- dinov3
- vision
- body-fat-estimation
- onnx
- pytorch
- computer-vision
- healthcare
- edge-ai
- sports-science
datasets:
- MasterMIARFID/gold_label_male_torso_dataset
- Sariax/MasterMIARFID_torsobf_dexa
base_model:
- facebook/dinov3-vitl16-pretrain-lvd1689m
extra_gated_heading: "Access Request — Non-Commercial Research & Educational Use Only"
extra_gated_prompt: "This model is available strictly for non-commercial academic, research, and personal educational purposes. Any commercial use, redistribution, SaaS/mobile app integration, or serving predictions via commercial APIs requires an official commercial license from the author."
extra_gated_fields:
Full Name: text
Institution, University, or Company: text
Official / Corporate Email: text
Intended Use (Academic Research / Clinical Study / Commercial Licensing Inquiry): text
I agree to use this model strictly for non-commercial purposes and not to redistribute weights: checkbox
---
# 🧬 DINOv3 ViT-L/16 — Whole Body Fat Percentage (WBFP) Estimator
-orange.svg?style=for-the-badge)
-red.svg?style=for-the-badge)
---
## 📌 Executive Summary
This repository contains the production-grade **Vision Transformer model (`39C / 39CC`)** developed as part of a Master's Thesis (TFM) at **Universitat Politècnica de València (UPV)** for the *Master's Degree in Artificial Intelligence, Pattern Recognition, and Digital Imaging (MIARFID)*.
The model automatically estimates **Whole Body Fat Percentage (WBFP)** from a **single frontal RGB torso photograph** completely **offline on-device (Edge AI)**, achieving clinical agreement with hospital **Dual-Energy X-ray Absorptiometry (DEXA)** gold standards without sending biometric imagery to cloud servers (100% GDPR / HIPAA compliant by design).
* **Author:** Sergi Cortés Guerrero (Universitat Politècnica de València)
* **Academic Tutor:** Dr. Juan Jesús Izquierdo Doménech (DSIC / ETSINF)
* **Gold Standard Reference:** Hospital DEXA Whole Body Scanners
* **Test Cohort (Frozen Hold-Out):** $N = 23$ patients with clinical DEXA ground truth
---
## 🔬 Clinical Evidence & Validation (DEXA Ground Truth)
The model was rigorously validated on a completely isolated, frozen clinical hold-out test set ($N=23$) with ground-truth DEXA measurements:
### 1. Correlation & Bland-Altman Agreement vs DEXA
As shown in the clinical validation plot above:
* **Pearson Correlation Coefficient ($r$):** **$0.92$** ($p < 0.001$, $95\%\text{ CI } [0.80, 0.98]$)
* **Standard Error of Estimate (SEE / RMSE):** **$2.96\%$** (with 10x Test-Time Augmentation) / **$2.98\%$** (standard single-pass)
* **Mean Absolute Error (MAE):** **$2.20\%$** (with 10x TTA) / **$2.28\%$** (standard single-pass)
* **Bland-Altman 95% Limits of Agreement:** Zero systematic proportional bias across the entire physiological spectrum (from lean athletic individuals at $6\%$ to clinical obesity at $>38\%$).
---
### 2. Research Progression & Architectural Breakthroughs
Across 119 controlled experiments, Vision Transformers fine-tuned with Low-Rank Adaptation (LoRA) and optimized via Knowledge Distillation significantly outperformed traditional ConvNets, 3D parametric morphing (SMPL), and Multimodal Large Language Models (VLLMs):
| Architecture / Paradigm | Description | Test SEE (%) | MAE (%) | Pearson $r$ | Clinical Assessment |
| :--- | :--- | :---: | :---: | :---: | :--- |
| **SMPL 3D Mesh + SVR** | 3D Body Mesh Morphing | 7.81% | 6.20% | 0.58 | Insufficient surface texture details |
| **ConvNeXt-Base** | Deep Convolutional Baseline | 6.84% | 5.21% | 0.72 | Severe overfitting on small cohorts |
| **Gemini 3.1 Flash (VLLM)** | Zero-shot visual prompt | 4.92% | 3.85% | 0.82 | High variance at physiological extremes |
| **DINOv3 ViT-L/16 Base** | Direct Scalar MSE Regression | 4.20% | 3.30% | 0.84 | Baseline self-supervised ViT |
| **Model 35-A (Distributional)** | 41-Bin Wasserstein Expected Loss | 3.75% | 2.85% | 0.88 | Captures epistemic visual uncertainty |
| **Model 37-DA (Silver Curated)** | Source-Differentiated LoRA | 3.23% | 2.45% | 0.90 | High precision on high-adiposity cohorts |
| **Model 39-C (Production Single)** | **Teacher-Distilled Monomodel** | **2.98%** | **2.28%** | **0.92** | **Production Single Model (This Repo)** |
| **Model 39-CC (Single + TTA 10x)** | **Teacher-Distilled + 10x TTA** | **2.96%** | **2.20%** | **0.92** | **Best On-Device Single Model SOTA** |
| **Ensemble HSR-3Way (Oracle)** | 3-Way Specialist Routing | **2.61%** | **1.98%** | **0.95** | **Global Benchmark Record (Enterprise Tier)** |
---
## 📱 Edge AI & Mobile Deployment (100% On-Device)
This model was specifically designed for **real-time on-device inference** in mobile applications without requiring internet access or cloud GPUs:
### On-Device Hardware Benchmark (Snapdragon 8 Gen 2 / Apple A16 Bionic)
| Weight Variant | Format | File Size | Size Reduction | Mobile CPU Latency | Accuracy Loss |
| :--- | :---: | :---: | :---: | :---: | :---: |
| `final_epoch_model.pth` | PyTorch FP32 | 1.21 GB | Baseline | ~580 ms | 0.00% (Baseline) |
| `final_epoch_model_flat.onnx` | ONNX FP32 | 1.16 GB | -4.1% | ~540 ms | 0.00% (Lossless) |
| `final_epoch_model_fp16.onnx` | ONNX FP16 | 581 MB | **-51.5%** | ~310 ms | < 0.01% (Undetectable) |
| `final_epoch_model_selective_int8.onnx` | ONNX INT8 | **294 MB** | **-75.6%** | **~210 ms** | **< 0.05% (Clinical Grade)** |
---
## 📦 Available Files & Checkpoints
All model variants are accessible upon approved gated access:
* `final_epoch_model_flat.onnx` (`1.16 GB`): Self-contained single-file FP32 ONNX model compatible with `onnxruntime` on Android, iOS, Windows, Linux, and macOS.
* `final_epoch_model_fp16.onnx` (`581 MB`): Half-precision FP16 ONNX model optimized for mobile GPUs and NPUs.
* `final_epoch_model_selective_int8.onnx` (`294 MB`): Dynamic INT8 quantized ViT backbone with FP32 head, designed for low-memory mobile CPUs.
* `final_epoch_model.pth` (`1.21 GB`): Native PyTorch FP32 weights (DINOv3 ViT-L/16 backbone + LoRA adapters + distributional head).
---
## 💻 Python Inference Example
```python
import numpy as np
import onnxruntime as ort
from PIL import Image
from huggingface_hub import hf_hub_download
# 1. Download model from gated HF repository (requires approved access token)
model_path = hf_hub_download(
repo_id="Sariax/dinov3-body-fat-onnx",
filename="final_epoch_model_flat.onnx",
token="YOUR_HF_TOKEN"
)
# 2. Initialize ONNX Session (CPU or CUDA)
session = ort.InferenceSession(model_path)
# Bin centers from 5.0% to 45.0% (41 continuous bins)
bin_centers = np.linspace(5.0, 45.0, 41)
# 3. Preprocess Image (224x224, ImageNet Normalized NCHW)
def preprocess(image_path):
img = Image.open(image_path).convert('RGB').resize((224, 224))
arr = np.array(img).astype(np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
norm = (arr - mean) / std
tensor = np.transpose(norm, (2, 0, 1)) # HWC -> CHW
return np.expand_dims(tensor, axis=0) # Add batch dim -> [1, 3, 224, 224]
# 4. Predict Softmax Distribution
input_tensor = preprocess("torso_photo.jpg")
logits = session.run(None, {"input": input_tensor})[0]
probabilities = np.exp(logits - np.max(logits)) / np.sum(np.exp(logits - np.max(logits)))
# 5. Compute Mathematical Expectation (WBFP)
estimated_wbfp = np.sum(probabilities * bin_centers)
uncertainty_std = np.sqrt(np.sum(probabilities * ((bin_centers - estimated_wbfp) ** 2)))
print(f"Estimated Body Fat: {estimated_wbfp:.2f}% (Visual Uncertainty: ±{uncertainty_std:.2f}%)")
```
---
## 💼 Commercial Licensing, Partnerships & Consulting
### Licensing Overview
| Tier | Target Audience | Permitted Use | Access |
| :--- | :--- | :--- | :--- |
| **Academic & Personal** | Students, Academic Researchers | Non-commercial research, thesis validation, open science | **Free** via Gated Access (CC BY-NC-ND 4.0) |
| **Commercial Startup / App** | Fitness Apps, Telemedicine, Smart Scales | Mobile/Web apps, SaaS products, freemium services | **Commercial License Agreement** |
| **Enterprise Clinical Tier** | Hospitals, Insurances, Healthtech Giants | Full **Oracle Ensemble (SEE 2.61%)** + Cloud / Edge deployment | **Enterprise Partnership** |
### Prohibited Without Commercial Agreement
Any commercial deployment, bundling into paid or ad-supported apps, serving predictions through commercial APIs, or redistribution of weights without a signed commercial agreement is strictly prohibited under the **Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)** license.
### Commercial Inquiries & Custom Development
If you or your company wish to:
1. **Acquire a commercial license** for the on-device production model.
2. **Access the Oracle Ensemble 3-Way HSR system** (benchmark record: **SEE 2.61%, MAE 1.98%, Pearson $r = 0.95$**).
3. **Commission custom engineering:** Mobile app integration (Flutter, Swift, Kotlin), camera-guided alignment, calibration on female/athletic cohorts, or enterprise cloud deployment.
👉 **How to contact:**
* **Hugging Face:** Open a formal inquiry in the [Discussions tab](https://huggingface.co/Sariax/dinov3-body-fat-onnx/discussions).
* **Direct Contact:** Contact the author **Sergi Cortés Guerrero** directly via Hugging Face profile or through the contact details provided upon gated access approval.
---
## 📚 Citation
If you use this model or refer to its methodology in academic work, please cite:
```bibtex
@mastersthesis{cortes2026bodyfat,
title = {Estimaci{\'o}n del porcentaje de grasa corporal mediante modelos de visi{\'o}n multimodal y despliegue en dispositivos m{\'o}viles},
author = {Cort{\'e}s Guerrero, Sergi},
school = {Universitat Polit{\`e}cnica de Val{\`e}ncia (UPV)},
year = {2026},
type = {Master's Thesis},
address = {Valencia, Spain}
}
```