This workflow uses Z Image Base to generate a strong, accurate base image with high prompt adherence at low CFG values.
Generate the base image at around 20 steps using Z Image Base.
This stage focuses on structure, composition, and correctly following complex prompts.Apply Z Image Turbo as a latent detailer.
The refinement happens entirely in latent space, so the image is not decoded or resized during this step.
This avoids quality loss and keeps the original structure intact.Turbo Detailer enhances:
Micro details like hair strands, sparks, fur, and surface deformation
Facial realism and texture
Recovery of realism from painterly or anime-style base outputs
Finish with Seed VR upscale for native-resolution upscaling, producing a clean, high-definition final image.
This method balances prompt accuracy, detail refinement, and image quality, making it effective for both realistic and semi-stylized images.
Latent Upscale (Turbo Detailer) – Recommended Settings
After generating the base image, the refinement is done using latent upscaling with Z Image Turbo.
Latent Upscale Method
Upscale in latent space
Scale: 1.5×
Upscale method: Bicubic
No decoded image resize before refinement
Turbo Detailer Sampling
Steps: 3
CFG: 1.0
Sampler: Euler or DEIS
Scheduler: SGM Uniform or Beta
Denoise:
First latent refine pass: 0.50
Final refinement pass: 0.15
Why these values
Low step count keeps the structure stable
Low CFG avoids overcooking details
Higher denoise (0.50) allows Turbo to add structure-level detail
Final low denoise (0.15) sharpens micro-details without altering composition
This approach enhances fine texture, realism, and detail density while preserving the original prompt intent and composition.