Sign In

Kimji

0

Download

1 variant available

SafeTensor

162.25 MB

Verified:

Type
LoRA
Stats

66

0

Reviews
Published

Aug 16, 2026

Base Model

ZImageTurbo

Training
Steps: 2,000
Epochs: 10
Hash
AutoV2
5E75367BA8
Trigger Words
kimji
default creator card background decoration
Followers - 33

33

Likes - 20

20

Created on Civitai

License:

Apache 2.0
XJGFRBX058X6KNSB9D3CMEV210.jpeg

Kimji – Z-Image Turbo Character LoRA

Kimji is a character/identity LoRA trained for Z-Image Turbo (ZIT), designed to maintain a consistent adult East Asian female character across different camera angles, compositions, expressions, outfits, lighting conditions, and environments.

The training dataset contains multiple facial angles, including frontal, three-quarter, and side-profile views, together with variations in composition, expression, hairstyle, clothing, lighting, and background. The goal is to preserve Kimji's recognizable facial identity while retaining enough flexibility for prompts to control the scene, pose, styling, and photographic appearance.

Trigger Word

kimji

Use kimji in the positive prompt when generating the character.

  • Base Model: Z-Image Turbo

  • Recommended Version: Epoch 9

  • Recommended LoRA Strength: 0.65–0.80

  • Suggested Starting Strength: 0.70

Testing showed that Epoch 9 at approximately 0.70 strength provides a good balance between facial identity and generation flexibility.

Epoch 10 can provide slightly stronger identity influence, while Epoch 8 remains a useful alternative when a softer LoRA influence is preferred.

Usage

Simple starting example:

kimji, adult East Asian woman, photorealistic portrait, natural skin texture, realistic photography

Then describe the desired expression, hairstyle, clothing, pose, camera angle, composition, lighting, and environment normally.

The LoRA is intended primarily to provide character identity, rather than permanently locking Kimji to a particular hairstyle, outfit, expression, pose, or background.

Notes

Results can vary depending on the Z-Image model/checkpoint, prompt, generation settings, and LoRA strength.

For most generations, start at 0.70 and adjust the LoRA strength in small increments.

If facial identity becomes too weak in complex scenes or unusual camera angles, try increasing the strength toward 0.75–0.80. If the LoRA begins to overpower the requested styling or scene characteristics, reduce the strength toward 0.60–0.65.

Trigger: kimji
Base Model: Z-Image Turbo
Recommended Epoch: 9
Recommended Strength: 0.65–0.80
Suggested Starting Point: 0.70