Z-Image

Z-Image Turbo Upscaler + Detailer ComfyUI Workflow: Complete Guide

May 13, 2026

Z-Image Turbo Upscaler + Detailer ComfyUI Workflow: Complete Guide

Published: 2026-05-13
Author: Z-Image Tech Team
Tags: Z-Image Turbo, ComfyUI, Upscaler, Detailer, Workflow, Image Upscaling

Overview

Z-Image Turbo is renowned for its blazing-fast 8-step image generation, but its native output resolution is typically capped at 1024×1024. When you need higher-resolution, more detailed images, the Upscaler + Detailer two-stage workflow is the optimal solution.

This guide walks you through building a complete ComfyUI workflow that requires only 4 model files to achieve a full pipeline from standard output to 4K high-definition images.

Workflow Architecture

The complete Upscaler + Detailer workflow consists of three stages:

Stage 1: Base Generation (Text-to-Image)

Generate the base image using Z-Image Turbo:

  • Model: z_image_turbo.safetensors
  • Sampler: Euler
  • Steps: 8-12
  • CFG: 1.5-2.0
  • Output Resolution: 1024×1024

Stage 2: Super-Resolution Upscaling (Upscaler)

Scale up the base image by 2-4x:

  • Recommended Upscale Models:
    • 4x-UltraSharp.pth — Best general-purpose upscaler
    • 4xFaceUpDAT.pth — Face-specialized, accurate skin tone
    • RealESRGAN-x4plus.pth — Classic choice, fast processing
  • Scale Factor: 2x (recommended) or 4x
  • Output Resolution: 2048×2048 or 4096×4096

Stage 3: Detail Refinement (Detailer)

Use Z-Image Turbo to enhance details on the upscaled image:

  • Model: z_image_turbo.safetensors (same as Stage 1)
  • Denoise: 0.15-0.35 (low preserves structure, high adds detail)
  • Steps: 8-12
  • CFG: 1.5-2.0
  • Key: Provide detailed refinement prompts

ComfyUI Node Configuration

Core Node List

1. CheckpointLoaderSimple — Load Z-Image Turbo model
2. CLIPTextEncode (Positive) — Positive prompt
3. CLIPTextEncode (Negative) — Negative prompt
4. EmptyLatentImage — Initial latent (Stage 1)
5. KSampler — Base generation sampling
6. VAEDecode — Decode to pixels
7. ImageUpscaleWithModel — Super-resolution upscaling
8. UpscaleModelLoader — Load upscale model
9. VAEEncode — Re-encode to latent (Detailer stage)
10. KSampler — Detail refinement sampling
11. VAEDecode — Final decode
12. SaveImage — Output

Node Connection Flow

CheckpointLoaderSimple → CLIPTextEncode(+/-)
                              ↓
                          KSampler(Stage 1)
                              ↓
                          VAEDecode
                              ↓
                     +-------+-------+
                     |               |
                SaveImage(original) ImageUpscaleWithModel
                                         ↓
                                    VAEEncode
                                         ↓
                                    KSampler(Stage 2)
                                         ↓
                                    VAEDecode
                                         ↓
                                    SaveImage(final)

Detailed Parameter Tuning

Upscaling Stage Parameters

Parameter Recommended Description
Upscale Model 4x-UltraSharp Best for general scenarios
Face-Specific 4xFaceUpDAT Portrait priority
Scale Factor 2x Balance quality and speed
Tile Mode Off Z-Image Turbo natively supports large images

Detailer Stage Parameters

Parameter Recommended Description
Denoise 0.25 Sweet spot, balances structure and detail
Steps 8-12 Z-Image Turbo optimal range
CFG 1.8 Medium guidance strength
Detail Prompt See below Enhance specific areas

Detail Refinement Prompt Templates

Portrait Refinement:

extreme detail, sharp eyes, skin texture, realistic pores, fine hair strands,
professional photography, 8K UHD, highly detailed

Product Refinement:

product photography, sharp focus, studio lighting, clean background,
high resolution, professional product shot, reflective surfaces,
fine texture details

Landscape Refinement:

landscape photography, ultra detailed, atmospheric perspective,
sharp focus, professional camera, 8K, HDR, fine detail

Practical Workflow Templates

Template 1: Quick Portrait Refinement

Stage 1: Generate
- Resolution: 1024×1024
- Steps: 8
- CFG: 1.5

Upscale:
- Model: 4xFaceUpDAT
- Factor: 2x

Stage 2: Refine
- Denoise: 0.20
- Steps: 10
- Prompt: extreme facial detail, sharp eyes, skin texture, professional portrait

Template 2: Product Photography Upscale

Stage 1: Generate
- Resolution: 1024×1024
- Steps: 10
- CFG: 2.0

Upscale:
- Model: 4x-UltraSharp
- Factor: 4x

Stage 2: Refine
- Denoise: 0.30
- Steps: 12
- Prompt: product photography, studio lighting, sharp focus, 8K

Template 3: ControlNet-Assisted Refinement

For scenarios requiring structural consistency, add a ControlNet Tile node in the Detailer stage:

Upscaled Image → ControlNet Tile Preprocessor → ControlNet Tile Model
                                                         ↓
                                                   KSampler(Stage 2)

Advantages:

  • Prevents structural drift during the Detailer stage
  • Preserves original composition
  • Ideal for architecture, products, and precision-demanding scenes

Common Issues and Solutions

Issue 1: Blurry After Upscaling

Cause: Mismatched upscale model or denoise value too low.

Solution:

  1. Try 4x-UltraSharp instead of RealESRGAN
  2. Increase denoise from 0.15 to 0.25
  3. Add sharp, crisp, detailed to detail prompts

Issue 2: Over-Modification in Detailer Stage

Cause: Denoise value too high, model "over-creates."

Solution:

  1. Lower denoise to 0.15-0.20
  2. Use ControlNet Tile for constraint
  3. Keep positive prompts concise

Issue 3: Out of Memory

Cause: 4K output requires significant VRAM.

Solution:

  1. Use GGUF Q4 quantized models
  2. Process in tiles (tile size: 512, overlap: 64)
  3. Two-step upscaling: 2x then 2x, rather than 4x at once

Performance Benchmarks

Configuration Generation Upscaling Refinement Total VRAM
RTX 3090, 2x ~4s ~3s ~5s ~12s ~8GB
RTX 3090, 4x ~4s ~6s ~8s ~18s ~12GB
M4 Mac, 2x ~8s ~6s ~10s ~24s ~16GB

Advanced Techniques

1. Multi-Stage Progressive Upscaling

For ultimate quality, use a three-stage workflow:

1024px → (2x) → 2048px → (refine) → 2048px → (2x) → 4096px → (refine) → 4096px

2. Face Detailer Dedicated Node

The ComfyUI community's Impact Pack provides a FaceDetailer node that automatically detects faces and performs local refinement:

Upscaled Image → FaceDetailer → Output

Parameters:

  • Detection threshold: 0.5
  • Detailer denoise: 0.3
  • Face mesh: Default

3. Batch Processing

Use ComfyUI's Batch node with folder input for batch upscaling and refinement:

FolderInput → ImageUpscaleWithModel → VAEEncode → KSampler → VAEDecode → SaveImage

Summary

The Z-Image Turbo Upscaler + Detailer workflow is a powerful tool for improving image quality and resolution:

  • Only 4 model files needed, simple setup
  • Fully automated end-to-end, from generation to output
  • 8K output, meeting professional-grade requirements
  • Multi-scenario adaptation, with templates for portraits, products, and landscapes

Master this workflow and your Z-Image Turbo generation capabilities will reach a whole new level.

Z-Image Team

Z-Image Turbo Upscaler + Detailer ComfyUI Workflow: Complete Guide | Blog