CKPT LoRA for Z-Image Base

LD0.82 LoRA

This is a LoRA extracted from an anime-style model. While LD0.82 shares similar tendencies with LD0.81, it differs slightly; in a sense, it represents the kind of anime model commonly encountered in image generation.

As before, I have prepared two versions: LD82.safetensors and LD82_detail.safetensors. Since LD0.82 is based on LD0.81, the LoRA represents the difference relative to LD0.81. The scaling factors are 1.5x and 4x (for the detail version). Because only up to LD0.80 has been released as a full checkpoint (ckpt), if you use LD0.80 as the base, applying LD081a and LD82 at a weight of 0.67 each, along with the detail version at 0.25, will yield results close to LD0.82 (though in practice, due to some loss, applying slightly higher weights will bring the result closer).

You can also create a wide variety of outputs by combining it with LD81a.

LD0.81 LoRA

While LucidDreamer ver.0.80 was created by reverting to an earlier model and rebuilding it, the overall aesthetic leaned towards a "cool" style. Some users might find a more standard "cute" style easier to work with; that is why I created this LoRA. Although technically a LoRA, it was effectively produced by taking LucidDreamer ver.0.80 as a base, fine-tuning it to match the characteristics of ver.0.81, and then converting the difference into a LoRA file.

The resulting images generally have a "cute" aesthetic, and the cel-shaded coloring becomes slightly softer. Since it only alters the style, there aren't many added elements, though it may perform better with portrait (vertical) or landscape (horizontal) aspect ratios.

LD81a.safetensors

This file captures the primary features using a rank of 128. It handles the broad stylistic elements.

LD81a_detail.safetensors

This file captures the details missed by the main LoRA using a rank of 64. It handles the finer stylistic nuances.

Both are intended for use at a weight of around 1.0, but they have been scaled higher than the actual values ??(1.5 and 4.0, respectively). Please keep in mind that a setting of 1.0 might be slightly too strong. While using these two files won't perfectly reproduce the original checkpoint, they can replicate it to a significant degree. You can also compensate for missing details by using higher weights.

Since they were trained on LD0.80, that model is recommended, but they should work with most Z-Image Base models.