AI Tools
ComfyUI Face Detail Advanced Workflow — How to Use Z-Image Turbo
Generated an image with ComfyUI but unhappy with the face? Using an advanced face detail workflow leveraging Z-Image Turbo, you can meticulously refine just the facial area while maintaining consistency throughout the entire image. After reading this article, even beginners can get started without complex node configuration.
ComfyUI Face Detail Advanced Workflow — How to Use Z-Image Turbo
Z-Image Turbo is the most efficient AI image generation model as of 2026. Using only 6B parameters (neural network weights), it doesn’t require expensive graphics cards, allowing you to create professional-quality images on a regular PC. It’s particularly excellent for face detail refinement purposes.
ComfyUI is an AI image generation tool that lets you connect nodes (small work units) to perform complex tasks. A face detail workflow is an automated process that detects only the facial area of a generated image and regenerates that portion more precisely. It’s like selecting just the face in photo editing and applying corrections to it.
Why is this necessary? General image generation models struggle to render faces perfectly. This means eyes might be crooked, noses might look odd, or skin texture might appear unnatural. However, using Z-Image Turbo’s face detail workflow, you can keep the overall image unchanged and refine just the face a few more times to achieve much more natural and beautiful results.
ComfyUI Face Detail Advanced Workflow — Z-Image Turbo Usage Tutorial
Before starting, you need to prepare the necessary files and programs. Using the officially provided JSON workflow file allows you to start immediately with drag-drop method without complex node connections.
The simplest method is to use the workflow template provided in the official GitHub repository. For example, download the ‘image_z_image_turbo.json’ file and drag it into the ComfyUI window to automatically connect all nodes. This file can be found on the Z-Image official website or ComfyUI example pages.
After loading the template, follow these steps. First, enter the prompt (description of the image you want to generate) in the input field. Next, click ‘Generate Queue’ and then click the ‘Prompt’ button to start image generation. About 8 steps is sufficient, so generation speed is fast.
Basic structure for writing good prompts:
’30-year-old woman, long black hair, large eyes, bright smile, warm lighting, studio background, soft skin texture, soft focus, upper body portrait, professional photography’ — The more detailed you include age, character features, background, lighting, camera angle, atmosphere, texture, and style, the better the results.
ComfyUI Face Detail Advanced Workflow — Z-Image Turbo Setup Steps
To properly set up your Z-Image Turbo environment, you first need to download the required model files and save them in the correct folders. I’ll explain this process step by step.
- Install Latest ComfyUI Version — Click ‘Download ComfyUI’ on the official website (www.comfy.org). If you’re a Windows user, select ‘Download for Windows’ and proceed through installation with all default settings by clicking ‘NEXT’.
- Download Model Files — Download the following 3 files from the HuggingFace website. ① General Model: diffusion_models folder at huggingface.co/Comfy-Org/z_image_turbo ② General Text Encoder: text_encoders folder on the same site ③ VAE file: Handles image compression/decompression. If VRAM is insufficient, download the GGUF format file.
- Verify Correct Folder Structure — Save the downloaded files as follows. ComfyUI folder → Inside models folder, place ① text encoder file in CLIP folder ② VAE file in VAE folder ③ model file in unet folder respectively.
- Install Additional Tools — Install necessary extensions through ComfyUI Manager. Install ‘ComfyUI-GGUF’ (required when using GGUF files) and ‘REalESRGAN x4’ (tool for upscaling small resolutions to better quality). After installation, completely close ComfyUI and run it again.
- Load Official Workflow Template — When you run ComfyUI again, a blank screen appears. Drag the official JSON workflow file onto this screen and all nodes will automatically connect.
If you need to create nodes manually, hover over empty space and double-click to open the node selection menu. Required nodes are ‘CLIP Loader’ (for loading model text encoder) and ‘CLIP Text Encode’ (for writing prompts). Change the CLIP Loader type to ‘qwen_image’, then connect the CLIP output of CLIP Loader to the clip input of CLIP Text Encode.
ComfyUI Face Detail Advanced Workflow — Z-Image Turbo Workflow Tips
- ①Sampling Steps Setting — Since Z-Image Turbo is a distilled (optimized) model, 8-9 steps is recommended. Excessive steps (20 or more) can actually degrade quality.
- ②Adjust Guidance CFG Scale — Good quality can be achieved at default value of 0 or low values (0.5-1.0). High values (7.0 or more) can become unnatural by over-adhering to the prompt.
- ③Resolution and VRAM Optimization — 1024×1024px is recommended. If VRAM is insufficient, reduce to 768×768px or use FP8 quantized version. FP8 quantized format is essential for 8GB VRAM or less.
- ④Advanced Control with ControlNet — Download the Z-Image-Turbo-Fun-ControlNet-Union.safetensors file to enable edge detection, pose guide, and depth map-based generation. This allows more precise control over human pose and hand position.
- ⑤Include Sufficient Detail in Prompts — The more information you input, such as skin texture, lighting direction, emotional expression, and background atmosphere, the more faithfully results match the prompt.
Mac M1/M2 users can also run Z-Image Turbo, but generation speed is slow (approximately 4 minutes per 1024×1024 image). If you want faster generation, consider purchasing a graphics card like NVIDIA RTX 30 series (30GB) or higher.
ComfyUI Face Detail Advanced Workflow — Z-Image Turbo Important Notes and Warnings
If ComfyUI is not the latest version, Z-Image related nodes will not appear on screen. Be sure to update to the latest version before starting. If you don’t use the latest version as of 2026, you won’t be able to find the ‘Z-Image Turbo’ node itself.
If model files are not in the correct folder, the workflow will not work. CLIP files must be stored exactly in models/clip folder, VAE in models/vae folder, and unet model in models/unet folder. If placed in the wrong location, a ‘Missing node’ error will occur.
Korean text is not properly rendered. Even as of 2026, Z-Image Turbo handles English and Chinese well but Korean is imperfect. To generate Korean text (e.g., ‘woman drinking coffee at a cafe’), input in English or separately edit only the text portion from previously generated images.
Crashes due to insufficient VRAM — Generating at 1024×1024 resolution with less than 16GB VRAM may force the program to shut down. In this case, reduce resolution to 512×512 or 768×768, or use GGUF quantized models (Q3-K-M, Q4-K-M, etc.).
I’ve organized common issues encountered during generation and their solutions. If you encounter “CLIP text encoding failure” error, double-check that the model selected in CLIP Loader actually exists in the models/clip folder. The filename must be exact.
Another common mistake is leaving the negative prompt empty or writing it excessively long. Since Z-Image Turbo works well even with low guidance_scale values, start without a negative prompt first and add only “distorted face, blurry, low quality” if needed.
Finally, it’s good to start with simple prompts on first run. For example, test with something like “beautiful woman, portrait, soft lighting, 35mm lens” and gradually add more detail.
Start Right Now
Experience a new dimension of AI image generation with Z-Image Turbo. Following this guide step by step, anyone can achieve professional-level face detail refinement.