nadizik/synthetic-human-expressions-poses-3d

3D Synthetic Human Poses and FACS Expressions Dataset

This is a high-fidelity synthetic dataset consisting of 10,075 pairs of 3D human character renders and detailed natural language annotations.

Dataset Structure & Generation

To ensure consistency, the dataset is generated using a single base 3D human model. The diversity of the dataset is achieved through a wide range of body poses, facial expressions, and camera angles:

• Character: 1 base human model.
• Camera Setup: The camera rotates around the subject to capture the character from multiple viewpoints and angles.
• Variability: 10,075 unique combinations of distinct body poses and detailed facial expressions.

Unlike standard text-image datasets, this repository contains detailed 3D metadata (exact camera coordinates, focal lengths, and FACS-based facial blendshape intensities) along with validation diagnostics from a CLIP ViT model and an OpenCV ResNet-10 SSD face detector.

---

---

Dataset Preview

--- ## How to Download and Use

Since the dataset contains over 20,000 files, it is distributed as a single compressed archive for faster downloading and to prevent browser timeouts. Once extracted, the dataset occupies less than 100 MB of disk space

---

Option A: Automatic Loading via Python (Recommended - If using .zip format)

If you are using the .zip archive version, you can load and extract the dataset automatically using the Hugging Face datasets library:

pip install datasets

from datasets import load_dataset

# This will automatically download and extract the dataset structure
dataset = load_dataset("nadizik/synthetic-human-expressions-poses-3d", data_files="3D_DATASET_sorted.zip")

Option B: Manual Download & Extraction

Go to the Files and versions tab.

Download the 3D_DATASET_sorted.zip file.

Extract the archive using any standard tool (such as WinRAR or 7-Zip).
Once extracted, you will find the curated structure organized into four sub-directories based on visual alignment, CLIP scores, and face presence:

---
## Dataset Structure & Splits

The dataset has been strictly curated and categorized into four sub-directories based on visual alignment, CLIP scores, and face presence:

* **`with_face/` (5,285 pairs)**  
  High CLIP score ($\ge 0.25$) + Face successfully detected. Ideal for fine-grained facial expression training and mouth/eye tracking.
  
* **`without_face/` (2,287 pairs)**  
  High CLIP score ($\ge 0.25$) + No face detected + No facial/eye descriptors in the caption. Ideal for full-body posture, clothing, and action training.
  
* **`mismatch/` (1,238 pairs)**  
  High CLIP score ($\ge 0.25$) + Face not detected but mentioned in text (due to extreme camera angles, hand occlusions, or head twists).
  
* **`low_quality/` (1,265 pairs)**  
  CLIP text-image alignment score below the $0.25$ threshold (challenging or noisy correlation pairs).

---

### File Contents in Each Folder

Each category folder contains:
1. **`.jpg` ** — 3D rendered character images.
2. **`.txt`** — Corresponding text descriptions.
3. **`dataset_log.json`** — The original 3D generation log filtered specifically for the files in this category.
4. **`report_partial.json`** — A complete validation report recalculated specifically for the files in this category.

---

### JSON Metadata Structure Examples

#### 1. 3D Engine Metadata (`dataset_log.json`)
Each rendering maps to precise physical attributes in the engine log, containing exact physical cameras and standard Action Units (FACS):
```json
"view_000000.jpg": {
  "motion": "SMILE_BIG",
  "frame": 0,
  "total_frames": 12,
  "intensity": 0.5,
  "breath_value": 0.074,
  "breath_mode": "normal",
  "camera": {
    "x_translate": 0.0,
    "y_translate": 90.0,
    "z_translate": -400.0,
    "y_rotation": 180.0,
    "focal_length": 50.0
  },
  "properties": {
    "facs_ctrl_SmileFullFace": 0.246,
    "facs_ctrl_EyeLookSide-Side": 0.051,
    "facs_ctrl_EyeLookUp-Down": 0.148
  },
  "timestamp": "2026-06-16T22:35:06.280859"

2. Validation Metrics (report_partial.json)
The validation report tracks high-level statistical aggregates for dataset drift and quality monitoring:
code
JSON

download

content_copy

expand_less
{
  "threshold": 0.25,
  "total": 5285,
  "valid_count": 5285,
  "valid_percent": 100.0,
  "mean_score": 0.271,
  "std_score": 0.015,
  "low_score_count": 0,
  "face_stats": {
    "with_face": 5285,
    "without_face": 0,
    "method": "dnn"
  }
}
Intended Use Cases
This dataset is designed for:
Pose-Guided Generation (ControlNet / IP-Adapter): Training models on exact physical camera setups and body orientations.
Facial Expression Training (FACS Control): Fine-tuning text-to-image models to recognize Action Units (FACS) directly from descriptive text.
Multimodal Curation: Exploring the boundary where face-detection and semantic text-alignment diverge under extreme camera views.
License
This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Feel free to use it for both commercial and research purposes.