Z-Image

Z-Image is a powerful AI model with strong capabilities in photorealistic image generation, accurate rendering of both Chinese and English text, and robust adherence to bilingual instructions. It achieves performance comparable to or exceeding leading competitors with only 8 steps.

The Z-Image model adopts a Scalable Single-Stream DiT (S3-DiT) architecture. This design unifies the processing of various conditional inputs (like text and image embeddings) with the noisy image latents into a single sequence, which is then fed into the Transformer backbone. Text, visual semantic tokens, and image VAE tokens are concatenated at the sequence level to serve as a unified input stream, maximizing parameter efficiency compared to dual-stream approaches.

For a 6-billion parameter model, it performs exceptionally well in image generation. During testing on the ModelScope platform (which uses NVIDIA A10 GPUs), most generations took a maximum of only 2 seconds with just 9 steps. On high-end consumer GPUs (like an RTX 3090 or 4090), this would take roughly 2 to 3 seconds, while mid-range cards might take 4 to 5 seconds.

Prompt Enhancing & Reasoning: The powerful prompt enhancer (PE) uses a structured reasoning chain to inject logic and common sense, enabling the model to handle complex tasks like the 'chicken-and-rabbit problem' or visualizing classical Chinese poetry. In editing tasks, even when faced with ambiguous user instructions, the model can apply its reasoning capabilities to infer the underlying intent and ensure a logically coherent result.

Use cases

• Create photorealistic images with accurate bilingual text rendering in just 8 steps.
• Design bilingual posters with Chinese and English text.