Asian Tech Press (SEP 02) -- ByteDance Inc. recently open-sourced a compression technology codenamed OMGD, which is ByteDance's self-developed GAN (Generative Adversarial Network) compression algorithm, which can reduce the minimum arithmetic power consumption to 1/46 of the original while ensuring the same generation effect.

It is reported that the paper on the technique has been selected for the International Conference on Computer Vision ICCV 2021.

GAN is an essential deep learning model in artificial intelligence, widely used in image generation, music generation and video generation. It can also be used to improve image quality and achieve tasks such as image stylization and image coloring. For example, comic effects and other TikTok props are realized by GAN.

OMGD means online multi-granularity distillation. According to a paper by ByteDance's technical team, the algorithm can flexibly optimize and compress the GAN model during the training process, thus achieving better image results and less computational cost.

According to public reports, ByteDance has so far open-sourced projects such as Klever, a machine learning platform, Fedlearner, a high-performance distributed training framework, BytePS, and the LightSeq inference and training engine.