We have Released Neural Network Libraries v1.12.0! StyleGan2 and TecoGAN examples are now available!
StyleGan2 is a state-of-the-art model for image generation, with improved quality from the original StyleGan. The model introduces a new normalization scheme for generator, along with path length regularizer, both of which contribute to getting rid of artifacts present in the previous model. Currently, only inference code is available, but we plan to follow up quickly with training code.
We have also made Colab demo for StyleGan2. By changing a random seed fed to the network, you can see the generated faces change drastically. Not only simple face generation, you can try face generation with style mixing technique proposed in the paper as well (you can see that result below). Please give it a try and have fun!
TecoGAN (TEmporally COherent GAN) is a model for video super-resolution that exploits temporal self-supervision. By leveraging temporal adversarial learning with a novel ping-pong loss, the model is able to generate results free of temporal artifacts while preserving spatial details.
- Fix several problems that might cause crash or system resource use out.
- Extend watch dog time so not too sensitive
- Fix FusedBatchNormalization for ONNX Exporter.
- Rename variable for repeat node.
- Fill input roi of ONNX Resize.
- Add fixed-point support for more functions.