RedCraft & DarkBeast Z|Klein Exported LoRA Edition
2026.1.1 VX 新规执行，请移步TG： https://t.me/+Ka2NhoUF5W1iMDRl
Dark Beast Z8 SDA 🟥 ZITLoRA Edition 03/15/2026
DBZ 8 SDA 完整模型BF16：
https://civarchive.com/models/2242173?modelVersionId=2774410
DBZ 8 SDA workflow 配套工作流链接：
https://civarchive.com/models/579280?modelVersionId=2742099
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
Dark Beast KLEIN 9b 🟦 V2.0 BFS 03/03/2026
DBK v2 BFS最佳换脸工作流，配套模型：
https://civarchive.com/models/2242173
DBK BFS workflow 配套工作流链接：
https://civarchive.com/models/579280
This is the next-level face-swap specialized evolution of the Dark Beast lineage, built on the lightning-fast FLUX.2 Klein 9B accelerated model from Black Forest Labs.
Engineered with targeted optimizations for face swapping practices, it integrates BFS (Best Face Swap) technology to completely eliminate the rigid, unnatural look that plagued earlier face replacements — delivering seamless, lifelike integrations with preserved identity, expression, and lighting.
It also fully fixes the portrait reference issue from the previous DB BlitZ versions
Special thanks to the scheme provider: https://github.com/alisson-anjos for the powerful BFS foundation that powers this breakthrough.🟦
The author's In-site link: https://civarchive.com/user/NRDX
Important notes:
This version is exclusively designed around the Klein 9B accelerated edition — no base model exists.
Usage is identical to Black Forest Labs' official FLUX.2 Klein 9B accelerated release: ultra-low steps (e.g., 4-5), CFG=1 fixed, blazing inference speed on consumer hardware.
In one sentence: Dark Beast's ferocious soul meets BFS (Best Face Swap) technology — more natural, and truly unstoppable! 🟦
for more infomation about BFS (Best Face Swap) :
https://huggingface.co/Alissonerdx
Alternatively, it can be directly applied to the entire Klein 9b/Qwen Edit base and Fine-tune models, through LoRA Adapter parameter injection.
2024
RedCraft | 红潮 CADS Q4 GGUF Enhancer-LoRA
De-distill&CADS Commercial Materials
Support online generation of ComfyUI WebUI
matching：
RedCraft | 红潮 CADS Commercial & Advertising Design System
GGUF format Q4 quantization version
for 1024 x 1024 inference
better with This Pages ：RedCraft-Q4enhancer-LoRA
The LoRA weight setting range is from 0.2 to 0.65
Inference parameters:
10-20 STEPS Euler / DPM++2M | beta / SGM_Uniform
CFG 3-3.5
Real CFG needs to be set (ignore guidance, or set to 0)
Description
Dark Beast Z8 SDA 🟥 ZITLoRA Edition
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
---
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.
---
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.
新手司机慎用，该模型极易造成肾衰竭
---
[HF repo](https://huggingface.co/GuangyuanSD/REDCraft-DarkBeast-Z-Image-TURBO)