Z-Image Turbo - T2I WF + 6x Detailer + SDXL Definer + SeedVR2 Upscaler + Seed Variance Enhancer + NSFW Fix
Details
Download Files
About this version
Model description
β‘ Z-Image Turbo
Simple Text 2 Image Workflow + Upscaler & Lora Loader
Z-Image Turbo is a fast, distilled, powerful and highly efficient image generation model with 6B parameters. The strength of this model lies in its ability to generate photorealistic images. But even though the model follows prompts very well, the output still isnβt premium in terms of quality. We still need an upscaling process here to make the images presentable. This is where this workflow comes in, including the well-known Ultimate SD Upscaler.
π What do you need to get started right away?
This Workflow - just download it and load it into ComfyUI
Z-Image Model: z_image_turbo_bf16.safetensors (place in \models\diffusion_models\)
Text encoder file: qwen_3_4b.safetensors (place in \models\text_encoders\)
The VAE file: ae.safetensors (place in \models\vae\)
Various upscale models of your choice (place in \models\upscale_models\)
Regarding the upscale model I recommend 4x_NMKD-Siax_200k.pth or 4x_UltraSharp.pth. These deliver the best results. But feel free to try out different upscale models. You can find a good selection here.
If you used FLUX before then you already have the ae.safetensors file and dont need to download it again.
You can also use the fp8 checkpoint if you run a gpu with 8gb of vram. I see no reason to use GGUF models here as the BF16 and FP8 models are already small enough for every kind of GPU.
Make sure that you use the latest ComfyUI Core version. Also don't forget to update the ComfyUI Frontend package and your Custom Nodes. Otherwise compatibility problems may occur.
β Do I need some special nodes to run this workflow?
The nodes inside the workflow are common standard nodes that should be already installed on your ComfyUI. If ComfyUI still indicates a missing node just install it and restart briefly.
Custom nodes used in this workflow:
π οΈ Why the Text Encoder (Positive Prompt) is connected to negative and positive on the KSampler Node?
If the CFG value is set to 1.0 the negative prompt is ignored. This is one of the reasons why the model is so fast. This is intentional and not a mistake. Other workflows are ignoring this fact. Simply enter your positive prompt and you're good to go!
π‘ Usage Tips:
Simply set the desired resolution on the left, select your preferred upscaler on the far right and start generating.
If you only want to generate raw images for testing purposes, simply toggle off the upscaling group using the first node on the top-left.
Feel free to experiment with the denoise and cfg values ββin the ultimate sd upscale node to get the perfect result. Also try out different upscale models.
You don't need to go higher than 10 steps with this model. It is a 8 Steps Model. It will not generate better images if you go higher. Just leave it on 8, 9 or max 10 steps.
I like to have the option to directly select the aspect ratio (e.g. 5:7 or 9:16) in the node. If you prefer to enter the resolution manually simply disconnect or delete the "Base Resolution Node". Afterwards you can enter your desired resolution in the "latent image" node.
π v2 Changelog:
In comparison to my other results with SDXL I wasn't entirely satisfied. Therefore, I revised and optimized the workflow. I have made the following changes to v2:
- β¨ Added Toggle On/Off Node for Upscaling Group
Toggling the entire upscaler group just to generate RAW images was cumbersome. Now, a single on/off switch at the start lets you enable or disable the upscaler with just one click β quick and easy.
- β¨ Added Lora Loader Node
With more Lora's being released for Z-Image Turbo these days, this node lets you load and configure multiple Loras at once. Making experimentation faster and easier.
- β¨ Added Clear VRAM and Clear Cache Node
Placed between the Base Image Generation and Upscaler sections, these nodes improve memory management. Especially on GPUs with limited VRAM and slightly speed up the workflow.
- β¨ Added Seed Generator Node
The new Seed Generator node is connected to the KSampler and also to the Upscaler node. This means that the Upscaler receives the same seed. I feel this leads to better results.
- β¨ Added fast sharpen / unsharpen node
This node allows you to adjust the sharpness slightly after upscaling. Feel free to experiment with thestrengthvalue. However, I suggest leaving it at 0.50.
- β¨ Added fast film grain node
This node adds film grain after upscaling. This provides a professional finish and a cinematic look. The grain_intensity value can be increased up to 0.050. I find less is more here and therefore have it set to 0.010.
- β¨ Adjusted some parameters for better performance
I adjusted some parameters, mainly affecting the Image Upscale node. I found that 20 steps unnecessarily slowed down the upscaling process and resulted in a poorer outcome, especially regarding skin texture. Eight steps proved to be the perfect value. This reduced the overall workflow time to 30% while simultaneously improving the quality.
π οΈ All the changes combined now result in a much faster workflow and very decent images compared to v1.
If you have a question or comment about the workflow feel free to leave a post in the comment section.
Have fun and feel free to post your generated images in the gallery below! π
