metadata
license: mit
facegen-facenet-unet-gan-embedding
Model Overview
This repository hosts a generative model trained to synthesize 128×128 face images conditioned on facial embeddings. The architecture jointly trains a FaceNet encoder and a UNet-based GAN generator to produce high-fidelity images from identity embeddings.
Codebase, Dataset, and Model Artifacts
| Category | Description | Link |
|---|---|---|
| GitHub Repo | Full training and inference codebase | GitHub Repository |
| Dataset | Cropped faces (128x128) for training | Hugging Face Dataset |
| Trained Model | Final GAN model with FaceNet encoder | Hugging Face Model |
| Training Notebook | End-to-end model training pipeline in Colab | Colab Notebook |
| Inference Notebook | Generate images from embeddings | Colab Notebook |
Experimental Summary
Find detailed metrics, loss trends, and inference samples in the full Weights & Biases report:
Weights & Biases Report – Training, Metrics, and Sample Outputs
Use Case
The model can be used for:
- Face reconstruction from embeddings
- Conditional face generation
- Evaluating zero-shot generalization for unseen identities
Architecture
- Encoder: Pretrained FaceNet (jointly fine-tuned)
- Generator: Modified UNet with residual upsampling blocks and conditional embedding injection at multiple resolutions
- Embedding Injection: Done at 8×8 resolution with skip connections
Training Details
- Training Time: ~6 hours on A100
- Dataset: 12,000 human face images
- Losses: MSE + Perceptual + Adversarial
- Optimizers: AdamW with learning rate scheduling
Checkpoints
epoch_100/epoch_200/epoch_300/(final)
Intended Use
This model is intended for research purposes, particularly for understanding conditional generation and face embedding interpretability.
Limitations
- Trained on a limited dataset
- May not generalize well to non-human faces or distorted embeddings