Z-Image - Ali Tongyi Labs open source image generation model

What is Z-Image

Z-Image is an open source image generation model from Ali Tongyi Lab with efficient, fast and powerful image generation capabilities. It adopts the Single Stream Diffusion Transformer Architecture (S3-DiT), which integrates text, visual semantics and image VAE tokens into a unified input stream to maximize parameter efficiency. Its core technological innovations include Decoupled-Distribution Matching Distillation (DMD) and Reinforcement Learning and Distribution Matching Distillation Fusion (DMDR), which dramatically improve the performance and image quality of less-step generation.Z-Image-Turbo version can generate high-quality images with only 8 function evaluations, supports sub-second reasoning latency, adapts to low-memory devices, and excels in photo-realistic image generation and bilingual text rendering. The Z-Image-Edit version focuses on image editing tasks with precise editing based on natural language cues, and the Z-Image-Base is an undistilled base model that provides the community with a wider scope for fine-tuning and custom development.

Features of Z-Image

• Efficient and fast generation: The Z-Image-Turbo version generates high-quality images in as few as 8 function evaluations, realizes sub-second inference latency, and adapts to low graphics memory devices for rapid prototyping and creative exploration.
• Powerful text rendering: Supports bilingual text rendering, capable of accurately generating images that contain both Chinese and English, meeting the needs of multiple languages.
• Photo-realistic image generation: Specializes in generating images with natural lighting, realistic textures, and believable scenes that can be used for creative design and visual effects production.
• Creative Image Editing: The Z-Image-Edit version allows precise image editing based on natural language prompts and supports creative image-to-image generation for a wide range of creative needs.
• Open source and flexible applications: The code, weights and online demos are open source, follow the Apache 2.0 license, and can be widely used in commercial projects, providing rich customization and development space for developers.

Z-Image's Core Advantages

• Single Stream Diffusion Transformer Architecture (S3-DiT): Z-Image uses this architecture to stitch text, visual semantic tokens and image VAE tokens at the sequence level as a unified input stream to maximize parameter efficiency.
• Decoupled-DMD (Decoupled Distribution Matched Distillation): is the core step less distillation algorithm that empowers the 8-step Z-Image model. By decoupling the two mechanisms CFG Augmentation (CA) and Distribution Matching (DM), we study and optimize them independently, which significantly improves the performance of less-step generation.
• DMDR (Reinforcement Learning with Distribution Matching Distillation Fusion): Building on Decoupled-DMD, by synergistically integrating Reinforcement Learning (RL) with Distribution Matching Distillation (DMD) in the post-training phase of the less-steps model, semantic alignment, aesthetic quality, and structural consistency are further improved while generating images with richer high-frequency details.
• Efficient less-step reasoningDecoupled-Distributed Matching Distillation (Decoupled-DMD) technology is used to generate high-quality images in only 8 steps, with fast inference speed, suitable for low video memory devices, and low inference latency.
• Powerful text rendering: Supports bilingual text rendering in English and Chinese, and can accurately generate images containing complex text for multilingual environments.
• High quality image generation: Generate photo-realistic images with natural lighting, lifelike textures and believable scenes to meet the needs of demanding visual effects.
• Precision image editing: The Z-Image-Edit version allows precise image editing based on natural language commands and supports creative image-to-image generation with powerful editing capabilities.

What is the official website of Z-Image

• Project website:: https://tongyi-mai.github.io/Z-Image-blog/
• GitHub repository:: https://github.com/Tongyi-MAI/Z-Image
• HuggingFace Model Library:: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

Who Z-Image is for

• Creative Designer: It can quickly generate high-quality images to meet creative design needs and help designers quickly explore and realize creative ideas.
• content creator: Supports Chinese and English text rendering and image editing, suitable for producing visual content containing text, such as social media images, advertisement design, etc.
• Developers and researchers: The open source code and flexible architecture provide developers with rich customization and development space, suitable for secondary development and research and exploration.
• business user: Follow the Apache 2.0 license, can be applied to commercial projects, suitable for enterprises for product design, marketing material generation and other scenarios.
• Individual enthusiasts: Low video memory device adaptation and fast generation capabilities are easy to use by individual users, making it suitable for creative exploration by individual users interested in image generation.