Z-Image-Distilled Speeds up Z-Image in only 10 Steps

    
        By vramkickedin    
     | 
    
            March 17, 2026 at 2:06 pm        
    
     | 
    
        2 min read

Z-Image-Distilled is a new image generation model that speeds up the creation process while keeping the original Z-Image style intact. It produces quality images in just 10 to 20 steps, compared to the 28 to 50 steps typically needed by the base version.

Developer GuangyuanSD created this model specifically for testing LoRA training on the non-Turbo Z-Image base. The project targets users who need faster generation times without giving up the diversity and artistic freedom found in the original model.

Model Size: From 1.14 GB & VRAM GPU: requirements vary

Speed meets creative freedom

Generates images in 10–20 steps.
Retains the original Z-Image's style diverse randomness.
Offers strong LoRA compatibility with recommended weights of 0.6–1.0.
Works with ComfyUI using standard workflows.
Better fine-tuning potential.

Artists and designers who want to train custom LoRAs will find this model useful for rapid testing and iteration. It also suits small creative teams who are looking to generate concepts quickly.

Current issues for Z-Image-Distilled the developers plan on tackling

The current release has some known issues. The distillation process affects text rendering, making small text less clear than in the original model. Some samplers can also cause color cast problems, particularly an unwanted blue tint.

GuangyuanSD has outlined clear plans for future updates. The team aims to improve stability at very low step counts and enhance how the model handles negative prompts when using higher CFG values.