The design combines an encoder and task heads with a mixture-of-experts diffusion language-model backbone. Multi-codebook discrete visual representations encode each spatial position, with parallel processing across positions and sequential decoding across codebook levels. This connects visual generation and understanding within one modeling framework.
The project is aimed at multimodal researchers and developers exploring unified visual systems. The repository provides inference code and an Ascend-trained checkpoint. Its NVIDIA-trained checkpoint, full training code, and technical report are listed as later releases, while the project metrics are described as preliminary.
