Resources

Z-Image is a powerful and efficient image generation model with 6B parameters. Currently there are three variants: Z-Image-Turbo (distilled version, only 8-step inference), Z-Image-Base (base model) and Z-Image-Edit (image editing variant).

Experience Z-Image Turbo model on ModelScope platform, providing online inference service and API interface for quick integration and use by developers.

Z-Image Turbo model in HuggingFace community, providing complete model weights, usage examples and community support. Monthly downloads reach 111,244.

Z-Image De-Turbo is a de-distilled version of Tongyi-MAI/Z-Image-Turbo, fine-tuned on images generated by Z-Image-Turbo to break down the turbo distillation limitations. This model is specifically designed for training and deep fine-tuning, offering enhanced trainability and flexibility compared to the original turbo model.

This LoRA model enhances Z-Image's existing pixel art capabilities, making them more detailed and refined. No trigger words required, but using "pixel art" in prompts can achieve better results.

This is a ControlNet model with 6 blocks added, trained from scratch on 1 million high-quality image dataset for 10,000 steps, supporting multiple control conditions.

Z-Image Turbo official online demo application, maintained by Tongyi-MAI team, providing a platform to directly experience Z-Image generation capabilities. Runs on Zero GPU, no local configuration required.

Z-Image gallery application on ModelScope platform, showcasing Z-Image model generation effects in gallery format, providing Chinese interface optimization and integrated ModelScope ecosystem.

Smart frame application based on Z-Image, specially developed for MCP 1st birthday commemorative project, showcasing Z-Image's possibilities in creative application scenarios.

Z-Image custom LoRA model online gallery, hosting multiple custom-trained LoRA models including radical art styles and identity models, supporting online switching and preview functionality.

Interactive mystery game application developed based on Z-Image, combining image generation and gameplay, showcasing Z-Image's innovative use in entertainment application scenarios.

This paper proposes the DMDR framework, integrating reinforcement learning techniques into the distribution matching distillation process. Research shows that for reinforcement learning of few-step generators, the DMD loss itself is more effective than traditional regularization methods.

Z-Image project official homepage, providing comprehensive project introduction including core features, model architecture, performance evaluation and technical details. Supports bilingual content in Chinese and English.