None00000/CelebA_embedding_StyleGAN_to_VAE

🧠 Synthetic CelebA-StyleGAN Embedding Dataset

🧩 Dataset Creation Notes

Overview

This dataset was created by combining CelebA, StyleGAN (FFHQ), and Stable Diffusion (VAE Encoder) to generate facial embeddings annotated with visual attributes.

🔹 Step-by-step process

1. 1.Attribute Classifier (CelebA)
2. 2.Trained a ResNet-based classifier on the CelebA dataset to predict 40 facial attributes.
3. 3.Achieved ~92% accuracy on the validation set.
4. 4.The “Blurry” attribute was found to have no positive samples in the generated set, so only 39 valid attributes were retained in the final dataset.

1. 1.Synthetic Face Generation (StyleGAN FFHQ)
2. 2.Used StyleGAN trained on FFHQ to generate approximately 3.5 million synthetic face images.
3. 3.Each generated image was processed by the CelebA-trained classifier to obtain its 39-dimensional attribute vector.

1. 1.Embedding Extraction (Stable Diffusion VAE)
2. 2.Encoded each image using the VAE encoder from Stable Diffusion v1.5.
3. 3.The resulting embeddings represent each image in the Stable Diffusion latent space.

📦 Output Format

Each entry in the dataset includes:

• —`embedding` — latent vector from Stable Diffusion’s VAE
• —`attributes` — 40 binary attributes predicted by the CelebA-trained classifier
• —`source` — "synthetic" (generated by StyleGAN FFHQ)

⚠️ License and Intended Use

This dataset incorporates models and data under the following research licenses:

• —CelebA — Non-Commercial Research License
• —StyleGAN (FFHQ) — NVIDIA Research License
• —Stable Diffusion (VAE) — CreativeML OpenRAIL-M

Hence, this dataset is strictly for non-commercial research and educational purposes.