hanzi

Generate (fake) Chinese characters using GAN and VAE with tensorflow and keras.

Results

The resulting images are saved to *_samples/.

Training process

12MB gif (tanh in generator):

Some issues

• Using tanh in generator is more stable than Relu, and in some case no BN needed.
• In the generator, relu is much better than tanh in learning strokes.
• Relu is not suitable in the first layer of the generator.
• 判别器成中filter的数目决定了生成的字是否“残”。