zai-org/CogView2

Implements a three-stage hierarchical transformer (6B-9B-9B parameters) with custom local attention kernels for efficient token generation, featuring LoPAR acceleration and bidirectional completion via CogLM. Supports both text-to-image generation and text-guided inpainting with style control (photo, sketch, watercolor, etc.), optimized for A100 GPUs but scalable via batch size tuning. Built on SwissArmyTransformer framework with model hosting on Hugging Face Spaces and Replicate, primarily trained for Chinese/English text inputs.

955 stars. No commits in the last 6 months.