Z-image

Z-image is an open-source foundation model for text-to-image generation built around an efficient 6B-parameter Single-Stream Diffusion Transformer. Instead of scaling to extremely large parameter counts, Z-Image focuses on systematic optimization to reach top-tier output quality with far lower compute and memory requirements. The result is a model that can produce photography-level, photorealistic images while also handling challenging tasks such as rendering readable text in both Chinese and English.

Designed for practicality, Z-Image targets fast inference and strong performance per GPU. In typical “turbo” settings it can generate images in very few diffusion steps (for example, around 8 steps), enabling sub-second latency in favorable setups. The architecture and optimizations also make it more accessible to developers: it is intended to run on consumer-class GPUs with under 16GB of VRAM, lowering the barrier to integrating high-quality image generation into real products.

Beyond raw realism, Z-Image emphasizes semantic understanding and prompt alignment. It supports world knowledge and richer composition control so that prompts translate into coherent scenes, objects, and styles. It is also positioned as a creative tool: users can apply instruction following for edits that range from small local changes (e.g., modifying a specific object or region) to broader, global transformations (e.g., changing lighting, mood, or overall artistic style). more