zai-org/CogView4

Implements text-to-image generation using Diffusion Transformer (DiT) architecture with cascading diffusion and relay diffusion frameworks, supporting variable resolutions up to 2048×2048 pixels. CogView4 (6B parameters) uses GLM-4-9B text encoder with native bilingual support, while CogView3-Plus uses T5-XXL encoder optimized for English. Integrates with Hugging Face Diffusers pipeline and supports CPU offloading strategies to reduce memory footprint from 35GB to 13GB on consumer hardware.

1,106 stars. No commits in the last 6 months.