REDCraft-DarkBeast-Z-Image-TURBO

Model Structure and Generation Environment：ComfyUI

This is the RedCraft & Dark Beast Z-ImageTurbo (ZIT) FT's collection repo

Transferred from @AiMetatron CIVITAI model page：

RedCraft 」Z-Image 」红潮 & Dark Beast 」Z-Image Turbo 」黑兽

Welcome to subscribe My CIVITAI personal page , get more examples and workflows for free

---

RedCraft Z-Image Turbo 红潮 ZIT ⚡️

RED-Zimage Turbo 红潮造相 1.5 12/03/2025

RedZimage 版本介绍 1.5Version Introduction

Distilled from Nano Banana Pro big-data sampling from real user photos (thanks to 瓜哥@guahunyo),

based on the open-source research by 小智@xiaozhijason,synthesized into the dataset ZImageTurboGen-3k.

Improved the style of the ZiTurbo version that was heavily tuned on community AI-generated web images.

The dataset is specifically designed for the following research scenarios:

• Data distribution shift analysis 数据分布偏移分析

• Reversibility / de-distillation research of distilled models

0 蒸馏模型的可逆性/去蒸馏（de-distillation）研究

• Model diversity generation research 模型多样性生成

(HF): https://huggingface.co/datasets/lrzjason/ZImageTurboGen

The REDZ training set comes from the ultimate essence of the community

20+ year veterans' lightning-fast meme-battle handspeed.

(Thanks to 佬杨@wikeeyang, 亮总@亮亮rayne, 十字鱼@Gluttony100,可乐@Colour, and everyone else)

---

Dark Beast Z-Image Turbo 黑兽 ZIT 🟥

Dark Beast SDA 8 ZIT🟥Restoring Generative Diversity NSFW 03/15/2026

通过引用 @FOK 提出的新型 语义方向对齐（Semantic Directional Alignment, SDA）损失，本模型 merge 了 饼儿佬 训练的 LoKr 适配器，在完美保持 8 步基线模型极致锐度和快速推理速度的同时，恢复了 原始教师模型 70.2% 的组合多样性（LPIPS）

Designed to rescue the "Diversity Collapse" problem in few-step distilled Flow Matching / Diffusion models. By applying @FOK novel Semantic Directional Alignment (SDA) Loss, this Model adapter recovers 70.2% of the original teacher model's compositional diversity (LPIPS)

https://huggingface.co/F16/z-image-turbo-sda

https://modelscope.cn/models/FFFFFFoo/z-image-turbo-sda

DBZ 8 SDA Exported LoRa 版本模型 Rank256 FP16： https://civitai.com/models/964312?modelVersionId=2775165

DBZ 8 SDA workflow 配套工作流链接： https://civitai.com/models/579280?modelVersionId=2742099

✨经网友测试，V8 SDA 与 ZIT LoRA 存在兼容性问题（解锁自由度后无法对齐），正在与SDA作者沟通解决方案。

✨After testing by netizens, there is a compatibility issue between V8 SDA and ZIT LoRA (unable to align after unlocking degrees of freedom), and we are currently communicating with the SDA author for a solution.

---

Additional voice:

When distilling a 50-step diffusion model (teacher) into an 8-step model (student), the student tends to take shortcuts across the manifold and collapse toward the “mean” of the distribution. As a result, even when changing the random noise seed, the generated images are almost identical (with consistent character poses and layouts). Instead of directly matching the raw velocities, we fine-tune this model using a custom four-pillar architecture grounded in physical principles.

当将一个 50 步扩散模型（教师）蒸馏为 8 步模型（学生）时，学生模型往往会走捷径穿越流形，收敛到分布的“均值”。结果是，即使更改随机噪声种子，生成的图像也几乎完全相同（人物姿势、布局一致）。SDA方案没有直接匹配原始速度，而是使用一种基于物理原理的四支柱自定义架构对该模型进行微调（通过向量导航，老司机 FOK 的这套方法就像给流扩散模型装了“方向导航+安全带+聪明开关+最后刹车”，让它既敢大胆改构图，又不牺牲照片级清晰度。）

Through vector navigation, this approach is like equipping the flow diffusion model with "directional GPS + safety belt + smart switch + final brake," allowing it to boldly reshape compositions while preserving photo-level sharpness.

古有 Lvmin Zhang 控制网络锁姿势，今有 FOK佬 四条支柱破蒸馏！开源之基！🙏 LORD of OPEN DIFFUSION

Recommended Settings:

CFG Scale: 1

Steps: 8

Sampler & Scheduler: Euler + Simple

---

Key Features & Optimizations:

Restoring Diversity in 8-Steps

NSFW Unlock with 10% DPO color style and sharpness.

For more technical details and implementation of Diffusers, please visit:

@misc{sda_diversity_loss_2026,

title={Teacher-Guided Semantic Directional Alignment (SDA) for Restoring Diversity in Few-Step Distilled Models},

author={Fok},

year={2026},

url={https://huggingface.co/F16/z-image-turbo-sda}

}

---

Dark Beast BlitZ ZIT 🟥 黑兽 6 Beyond Reality NSFW 02/18/2026

Dark Beast 6 BlitZ ZIT based on SUPER Z V3.0 REBUILD (Vibrant Color Edition)

NSFW Unlocked Version: Z image turbo form @Nurburgring`s HF repo

Core Improvements:

Delivers superior color and lighting performance, with more refined texture details. Enhanced Capabilities: Supports a wider range of artistic styles compared to Z image turbo, while offering faster generation speeds than Z image base.

Recommended Settings:

CFG Scale: 1

Steps: 5～15

Sampler & Scheduler: Euler + Simple

Key Features & Optimizations:

Utilize nodes like SeedVarianceEnhancer to increase visual diversity. Optimized support for specific artistic styles and 2D Anime rendering.

🔧 LoRA Training Adapter Update

Important Notice: This is a LoRA Training Adapter specifically designed for the Beyond Reality 3.0 & Dark Beast 6 NSFW model.

Focus: Diffusion module optimization.

Repo Links:@Nurburgring`s HF repo

Concept: "Only new training brings new results; this is merely an adapter."

Compatibility: The Rebuild base model supports training on this adapter. You can replace it in aitoolkit to fine-tune your LoRA.

Created by @Laushine.

Usage Tip: Please change the file extension to ZIP before extracting and using it.

---

Dark Beast BlitZ ZIT 🟥 黑兽 5-steps 02/12/2026

In one sentence: Taking ZImage Turbo’s already blazing speed and cranking it to absolute BlitZ velocity while keeping every drop of that ferocious Dark Beast soul! 🟥

一句话：以 ZImage Turbo 已经炽热的速度， 将其调到绝对 BlitZ 速度，现在就释放出来！ 只需5步，同时保留那凶猛的黑暗野兽灵魂的每一滴！ 🟥

Key features:

Fully preserves the signature Dark Beast style, rich details, and intense Dark Beast aesthetic from the standard lineage

Refined through advanced targeted distillation & fine-tuning, now perfectly dialed in for zero-CFG guidance at minimal steps

BlitZ-level inference speed — breathtaking high-quality images in just 5 steps

Recommended settings: 5 steps, CFG=1 (fixed), any seed you want

Lightning-fast generation awaits — unleash it now! 🟥🚀