I have separated the anime-style and semi-realistic models. ImaginaryZ is a realistic-style model with a focus on sci-fi and fantasy themes.

During the development phase, the goal was to shift the output of the anime-oriented base model toward a realistic style while also aligning it with the output characteristics of Z-Image Turbo. I achieved this by training the model on Turbo-generated outputs and subsequently enhancing it with photographic datasets. Because this process involved a step akin to "winding back" the model, there are some areas where prompt interpretation may behave differently compared to LD-based models.

The model is designed to be used with a Distill LoRA. I recommend using Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors (the one used in the samples) at a slightly lower strength setting over approximately 15 steps; 8 steps are sufficient for previews.

ver.1.3HD

This is a high-resolution version that takes a different approach than ver.1.1HD. While version 1.1 tended to apply excessive shading, this version features a slightly more natural rendering style. Although it technically offers the highest resolution, the actual visual difference is minimal. It is not a matter of which version is superior to the others?including version 1.0. However, it does tend to exhibit significant color casting in indoor scenes.

ver.1.1HD

This is an FT version of ver.1.0 with slightly higher resolution data added. In addition, image quality is adjusted using LoRA. I'm trying to make it a little darker overall.