[Commercial Use Allowed] Z-Image_clear_photoreal Released | A Photorealistic Model Combining High Speed and High Quality [Image Generation AI]

Introduction

Hello, I'm Kimama / Easygoing.

This time, I am introducing the definitive version of an image generation AI that is fast, high-quality, and allows for commercial use: Z-Image_clear_photoreal.

What is Z-Image?

Z-Image is an image generation AI released by Alibaba Group in China on January 27, 2026, under the Apache-2.0 license, which also permits commercial use.

The Z-Image series gained popularity as an image generation AI that runs on mid-range PCs, following the release of the high-speed distilled version, Z-Image-Turbo, in November 2025.

While Z-Image-Turbo was a distilled model specialized for high-speed generation, the Z-Image (Base) released this time is the core model before distillation, which is essentially the raw brain of the trained AI.

Distilled models have less variety

Distilled models, including Z-Image-Turbo, undergo adjustments to improve usability.

Characteristics of distilled models

• Improved stability and speed

• Degraded image quality

• Reduced variety

Distilled models undergo adjustments to increase stability in order to reduce failures in illustration generation.

This allows illustrations to converge faster, enabling high-speed generation with low steps, but in exchange, vivid color expression is lost, the overall tone becomes brownish, and residual noise tends to remain in the finished illustration.

Furthermore, increasing stability makes it easier to generate similar illustrations, leading to a lack of variety in faces and compositions.

Mixing the core model based on the distilled model

For this Z-Image_clear_photoreal, we aimed to Z-Image-Turbo_clear, the distillation model introduced previously, as a base, and by mixing in the original Z-Image (Base) at specific layers, we aimed for improved image quality and variety.

By doing this, we aim for further improvements in image quality while maintaining the high speed without using CFG of Z-Image-Turbo.

Actual illustrations!

Now, let's compare the actual illustrations.

On the left is the newly adjusted Z-Image_clear_photoreal model, and on the right is the output from the original Z-Image-Turbo_clear.

Fireplace room

Late-night radio

Magnified view

When comparing image quality, it is easy to tell by looking at the skin texture; while the Z-Image-Turbo_clear on the right retains the granular noise characteristic of high-speed distillation models in the finished illustration, you can see that this has been improved and looks cleaner in the Z-Image_clear_photoreal on the left.

Comparing variety!

Next, let's compare the variety of the illustrations.

I will now generate eight consecutive illustrations while keeping the same prompt but changing the seed value.

Prompt (Woman in traditional attire)

Z-Image_clear_photoreal (This merged model)

Z-Image-Turbo_clear (Distillation model)

Comparing the two models, you can see that the newly adjusted Z-Image_clear_photoreal shows rich variety in the character's face and composition, whereas the original Z-Image-Turbo_clear has little change in faces or composition, resulting in similar illustrations.

Merge Recipe for Z-Image_clear_photoreal

Now, I will introduce the merge recipe for the Z-Image_clear_photoreal I created this time.

The Model Merge Z-Image node on the right is a node for mixing two Z-Image models, where the 0 setting value merges layers from Z-Image (Base), and the 1 setting value merges layers from Z-Image-Turbo_clear.

In other words, the initial layers use the diverse training data of Z-Image (Base) to increase variation, while the middle to later layers use the distilled Z-Image-Turbo_clear to maintain stability and generation speed.

Model Merge Z-Image Custom Node

Let's use Z-Image_clear_vae for the VAE!

For this Z-Image_clear_photoreal model, I recommend using it in conjunction with Z-Image_clear_vae, a custom model I created.

• Z-Image_natural_vae: Natural expression with suppressed saturation

• Z-Image_clear_vae: Vivid expression with increased saturation

Since I have prepared several Z-Image_clear_vae models with different color expressions, please try using them according to your preference.

Next time: Introduction to Z-Image_clear_anime!

This time, when adjusting Z-Image, the direction of adjustment was completely different between photorealistic illustrations and anime illustrations, so I created Z-Image_clear_photoreal as a dedicated model for photorealism.

Next time, I plan to introduce the Z-Image_clear_anime model for anime illustrations.

Summary: Let's try using Z-Image_clear_photoreal

• Z-Image (Base) has abundant variations

• Z-Image-Turbo offers stability and high-speed generation

• Z-Image_clear_photoreal takes the best of both worlds

This time, I introduced the Z-Image_clear_photoreal model.

When actually performing adjustments, I realized that Z-Image has almost no redundant structure, and it is a model that Alibaba has deeply researched Qwen-Image to achieve the best optimization and lightweighting.

With the release of the core Z-Image (Base) model, it has become easier than ever to develop the Z-Image series, and I believe that excellent improved versions from the community will emerge from here on.

Why not take this opportunity to try out Z-Image's high-quality, high-speed generation for yourself?

Thank you for reading until the end!

English Article