This repository documents the source code and training process for aConditional Variational Autoencoder Generative Adversarial Network Hybrid(cVAE-GAN) model used for the generative art project, I. by ryley-o.eth.
A small, efficient conditional cVAE-GAN implementation for 64x64 images using TensorFlow/Keras. This implementation is designed to be easily quantizable and deployable to blockchain via TensorFlow.js.
Note: LLM agent-based coding was used heavily when creating the code for this model. AI was embraced for reasons closely related to why this model was created.
- Conditional Variational Autoencoder (cVAE)
- Generative Adversarial Network (GAN)
- Conditional consistency and contrastive regularization
- Training schedule to blend cVAE's and GAN's strengths while focusing in interpolation stability
- Hyper-compact model
- Quantizable
- Deployable to blockchain
You can create the model that was uploaded to the blockchain by running the following commands:
# Train the model
python3 train_vaegan3_revb.py
# Export TFJS quantized model
python3 export_vaegan3_revb_tfjs.py
# Run python server
python3 -m http.server 8000
# Run the model in the browser, refresh for new images
open http://localhost:8000/simple_vaegan3_interp.html
# if you want to generate a grid of output images, run the following command
python3 generate_grid_vaegan3.py