CogView4 – Open-Source AI Text-to-Image Model Supporting Chinese Character Generation

What is CogView4?

CogView4 is an open-source text-to-image model developed by Zhipu AI, featuring 6 billion parameters and native support for Chinese input and text generation.

The model ranks first in overall scores on DPG-Bench, achieving state-of-the-art (SOTA) performance among open-source text-to-image models.

As the first image generation model released under the Apache 2.0 license, CogView4 supports arbitrary resolution image generation and can produce high-quality images from complex textual descriptions.

Key Features of CogView4

Bilingual Input Support – CogView4 is the first open-source model capable of generating Chinese characters, producing high-quality images from either Chinese or English prompts.

Arbitrary Resolution Image Generation – Supports resolutions ranging from 512×512 to 2048×2048, catering to diverse creative needs.

Strong Semantic Alignment – Ranks #1 on DPG-Bench, excelling in complex semantic alignment and instruction following.

Chinese Text Rendering – Optimized for Chinese text generation, allowing seamless integration of Chinese characters into images, making it ideal for advertising, short videos, and creative applications.

Memory Optimization & Efficient Inference – Features CPU offloading and quantized text encoders to reduce memory consumption and improve inference efficiency.

Technical Foundations of CogView4

Architecture – Combines diffusion models with Transformer architecture.

The diffusion model generates images by gradually removing noise.

The Transformer component handles joint representations of text and images.

Uses 6 billion parameters, supporting arbitrary-length text input and arbitrary-resolution image generation.

Text Encoder & Tokenizer –Uses a bilingual (Chinese-English) GLM-4 encoder for complex semantic alignment.

Converts text into embedded vectors via a Tokenizer, which are then integrated with the image’s latent representation.

Image Encoding & Decoding –Uses a Variational Auto-Encoder (VAE) to encode images into latent space representations.

The diffusion model then denoises the latent representations to generate the final image.

Diffusion & Denoising Process –The core mechanism of the diffusion model involves progressively refining images through denoising steps.

Uses FlowMatch EulerDiscrete Scheduler to control the denoising process.

Users can adjust inference steps (num_inference_steps) to balance quality and generation speed.

Multi-Stage Training Strategy –Includes base resolution training, multi-resolution training, high-quality data fine-tuning, and human preference alignment.

Ensures high-quality and aesthetically pleasing image generation.

Optimization & Efficiency –Implements memory optimization techniques such as CPU offloading and text encoder quantization to improve training and inference efficiency.

Released under Apache 2.0 license, enabling open-source development and community contributions.

Project Links

GitHub Repository: https://github.com/THUDM/CogView4

Hugging Face Model Hub: https://huggingface.co/THUDM/CogView4-6BZ

Use Cases of CogView4

Advertising & Creative Design – Seamlessly integrates Chinese and English text into images, generating high-quality posters, marketing visuals, and branding materials.

Educational Content Generation – Creates illustrations and scientific diagrams to help students understand complex concepts more effectively.

Children’s Book Illustration – Generates child-friendly artwork to inspire creativity and imagination.

E-commerce & Content Creation – Produces high-quality product images, advertisements, and promotional materials to help businesses attract customers.

Personalized Content Customization – Generates tailored visual content based on user-specific requirements, enhancing user experience.