Oct.29 2026, Moss version is updated:

Moss V1 Features：

MossImage SFT is an exceptional text-to-image model. However, after days of personal testing, I’ve identified several areas that could be optimized (purely from my own perspective):
For example, the lighting and shadow tend to look flat with insufficient contrast, high-frequency details appear blurry, human skin tones lean too red, colors are overly saturated, and scene elements — especially non-human ones — lack sufficient randomness (this issue becomes even more noticeable when using a speed-up LoRA).

To address these, I specifically trained the Moss-version Realism Aesthetic. After repeated testing on over 20 saved LoRAs, I selected this one to release as the V1 version.
Compared to the base SFT model, applying this LoRA can mitigate the above issues to a certain extent.

Key differences after using this LoRA

1. More realistic lighting and shadow, with improved contrast

2. Enhanced overall detail performance

3. Upgraded portrait aesthetic quality

4. Human skin tones no longer carry a strong red cast

5. Image color saturation becomes more natural

6. Higher scene realism with reduced "AI-generated" artifacts

You are all welcome to test it and share your output results. Hope you enjoy it!

Recommended Settings

• Weight: 0.8-1

• Sampler: Euler + Simple

• Steps: 25

• CFG Scale: 3.0

For now, using a speed-up LoRA alongside this one is not recommended.

MossImage SFT模型是非常优秀的文生图模型，但是经过这些天我个人的测试下来，我发现它还有很多可以优化的点（仅仅针对我个人来说）：

比如光影偏扁平，缺少对比度，高频细节模糊，人像的皮肤色调偏红，颜色过于饱和，场景元素尤其是非人物元素的随机性不够（用了加速lora后这个问题更明显）等。

因此，我特地训练了Moss版本的Realism Aesthetic，从保存的二十多个Lora中，经过我反复测试后，筛选出这个作为V1版本发布。

相比SFT模型本身，使用这个Lora后能一定程度上改善以上那些问题。

使用Lora后有何不同：

1.光影更真实，对比度提升。

2.细节提升。

3.人物美学提升。

4.人物皮肤色调不再那么偏红。

5.图像颜色的饱和度变得更自然了一些。

6.场景真实度得到提高，AI感降低。

欢迎大家测试返图，希望大家喜欢。

推荐设置：

权重：0.8-1，采样方式：Euler+Simple，25步，CFG=3.0

目前暂不建议使用加速Lora

Sep.20 2026, Krea2 version is updated:

Krea2 Features：

1. It can enhance the aesthetic appeal of characters, eliminate the distinct AI-generated look, and improve skin texture.

1. It makes landscapes, still lifes, indoor and outdoor scenes more photorealistic.

2. The lighting and shadows become more authentic, with better light transmission and clarity.

3. Richer and more details.

4. The overall aesthetic is significantly upgraded.

5. It features high generalization performance, does not alter the face of character LoRAs, and is fully compatible with both NSFW and SFW content.

‌Recommended weight‌: 1.

When used in combination with a photorealistic base model or multiple LoRAs, you can lower the weight to avoid overfitting.

‌Recommended sampling method‌: Euler+Simple, ‌steps‌: 10

This picture shows its effects on different Checkpoints.(Krea2RedMix1.1-INT8-Convrot-Native and Krea2-rayArtshoot_v5_int8_convrot)

Krea2版本特点：

1. 可以提升人物美感，去AI味道，提升皮肤质感。

2. 让景观、静物、室内室外等更写实。

3. 光影更加真实，通透感更好。

4. 细节增加。

5. 美学提升。

6. 高泛化性，不影响人物lora的脸型，对nsfw/sfw均兼容。

推荐权重：1，叠加写实大模型或多lora使用，可降低权重防止过拟合。

推荐采样：Euler+Simple， 步数：10

Aug.29 2026, ZIT version is updated:

Attention: ZIT lora weight can't be too high, 0.6 is recommend of this lora. Not recommend to using multiple loras because it can reduce the image quality, except to take a try on lower weight than 0.6. Lower weight than 0.6 may bring nothing to the image when using this lora, I think this is the deficiency of training on distilled model.

注意：ZIT的lora生图时，权重不要过高，经我测试推荐设置为0.6。 我不推荐该lora叠加其他更多lora一同使用，因为可能导致图像劣化，除非你自行测试降低权重，然而低权重可能会效果不明显，我感觉这些问题就是在蒸馏模型训练上训练lora的不足之处。

Recommended Settings:
Trigger Word: virealism
Weight:

• Z-Image Base LoRA used on Z-Image Turbo: 1.5

• Z-Image Base LoRA used on Z-Image Base: 0.7

• Z-Image Turbo LoRA: 0.6

• Krea2 LoRA: 1

ZIT comparison:

Here is the comparison between euler+ddim_uniform with euler+simple:

Aug.28 2026, ZIB version is updated:

I've long wanted to train a LoRA to reduce the artificial, AI-generated look of my images. To achieve this, I prepared a training dataset of real-world photography, and trained on Z-Image Base, Z-Image Turbo and Krea2. I will release these three LoRA models then. I found that while the performance of the LoRAs varies slightly across different base models, the following improvements are achieved overall:

• The overall lighting and shadow realism of the image is enhanced

• Human skin appears fairer and gains more translucency

• The contrast between light and dark areas is more pronounced

• Textures and details look more authentic

• The overall aesthetic presentation of human subjects is improved

• High generalization performance, no negative impact on the facial features of character LoRAs

Recommended Settings:
Trigger Word: virealism
Weight:

• Z-Image Base LoRA used on Z-Image Turbo: 1.5

• Z-Image Base LoRA used on Z-Image Base: 0.7

• Z-Image Turbo LoRA: 0.6

• Krea2 LoRA: 1

Sampler: Euler Scheduler: simple / ddim_uniform

Comparison/来对比一下看看：

我一直想训练个自己的lora来降低生图的AI感，于是我准备了真实摄影的训练集，在zimage base、turbo和krea2进行了训练，我会先后发布它们。我发现虽然在不同大模型上，lora效果表现略有差异，但总体得到了以下改善：

1. 画面整体光影真实感得到提升

2. 人物的皮肤更加白皙、通透感增加

3. 明暗对比呈现增强

4. 纹理与细节更真实

5. 人物美学提升

6. 高泛化性不影响人物lora面部

推荐设置：

触发词：virealism

权重：zimage base lora 在turbo上使用：1.5， zimage base：0.7

zimage turbo lora：0.6

krea2 lora：1

采样器：Euler 调度器：simple / ddim_uniform