Updated: Jan 31, 2026
tool
I just made a new Telegram Group, drop in for chats, questions, feedbacks or just to share your work. I will be sharing prompts and pics there as CIVITAI has been removing my work repeatedly.
V2.0 Dual Ksamlers
This is a classic dual Advanced KSampler build using ZIB + ZIT. Huge thanks to the AI artists in the group for sharing their experiments — based on that, I’ve reworked the flow to use a dual Advanced KSampler instead of the old two-pass KSampler setup.
The good stuff 🔥
V2 cuts generation time by roughly ~33%
Still hits strong prompt adherence while keeping most of ZIT’s speed
Image composition is still on par with V1
The trade-offs ⚖️
Preview is now a noisy latent only (V1 let you see rough image details on the first pass)
You’ll get a bit more ZIT and slightly less ZIB detail compared to V1, but ZIB’s composition strengths are still very much there
Also tossed in some misc fixes to clean up a few of my earlier mistakes 😅
这是一个经典的双 Advanced KSampler 架构，使用 ZIB + ZIT。非常感谢群里的 AI 艺术家们分享他们的实验成果——在此基础上，我重新调整了流程，改用双 Advanced KSampler，取代了旧的两次 KSampler方案。
亮点 🔥
V2 将生成时间缩短了大约 33%
在保持 ZIT 速度优势的同时，依然具备很强的提示词贴合度
图像构图效果仍然与 V1 持平
取舍 ⚖️
预览现在只显示带噪的 latent（V1 在第一遍就能看到大致的图像细节）
相比 V1，ZIT 成分稍多、ZIB 细节略少，但 ZIB 在构图方面的优势依然非常明显
另外还顺手修了一些杂项问题，清理了我之前的一些小失误 😅
V1.0 2 Ksampler Passes
All showcased result images are arranged as: ZIB only | ZIB + ZIT | ZIT only.
I’ve been experimenting with a new workflow lately. With the release of Zimage Base (ZIB), it introduces several advantages that Zimage Turbo (ZIT) doesn’t currently offer—most notably stronger prompt adherence, better compositional stability at low resolutions, and significantly higher variation when iterating with different seeds. These strengths make ZIB especially well-suited for defining the overall structure and layout of an image early in the pipeline.
However, until a properly optimized speed-up LoRA for ZIB is released, generation times remain relatively slow—unless you’re running on very high-end hardware. Because of that limitation, I developed a workaround:
I first generate a low-resolution pass using ZIB to quickly establish the base composition. This allows for faster iteration, broader variety across seeds, and more reliable prompt interpretation. Once the foundation is set, I switch to Zimage Turbo (ZIT) to refine and complete the finer details at higher resolution.
Why ZIB instead of SDXL, Pony, or other base models?
You can absolutely use those, but the key advantage of ZIB is ecosystem consistency. ZIB and ZIT can share the same CLIP, prompt structure, and VAE, which means everything integrates smoothly without needing to reload or re-import large components. This keeps the workflow lightweight, efficient, and cohesive.
The drawbacks are fairly obvious as well—ZIB cannot directly generate explicit (XXX) content. While private parts can be redrawn or enhanced during the second pass using specialized models like Moody Porn Mix, poses in those scenarios are still hit-or-miss. That said, with more XXX-focused LoRAs for ZIB expected in the near future, the lack of XXX poses should be resolved shortly.
This is an experimental workflow, and likely a temporary one until an official speed-up LoRA for ZIB is released—but it’s a practical and effective approach that’s absolutely worth trying in the meantime.
所有展示的结果分为三类：仅 ZIB | ZIB + ZIT | 仅 ZIT。
最近我一直在尝试一种新的工作流程。随着 Zimage Omni Base（ZIB）的发布，它带来了一些 Zimage Turbo（ZIT）目前尚未具备的优势——最显著的是更强的提示词遵循度、在低分辨率下更好的构图稳定性，以及在使用不同种子迭代时显著更高的变化多样性。这些优势使 ZIB 特别适合在流程早期用来确立图像的整体结构和布局。
然而，在针对 ZIB 发布真正优化好的加速 LoRA 之前，生成时间仍然相对较慢——除非你使用的是非常高端的硬件。正因如此，我设计了一个变通方案：
先用 ZIB 生成一轮低分辨率图像，快速确立基础构图。这样可以实现更快的迭代、在不同种子间获得更广泛的多样性，以及更可靠的提示词理解。一旦基础构图确定，我就切换到 Zimage Turbo（ZIT），在更高分辨率下精炼并完成细节。
为什么选择 ZIB 而不是 SDXL、Pony 或其他基础模型？
当然可以使用那些模型，但 ZIB 的核心优势在于生态系统一致性。ZIB 和 ZIT 可以共享相同的 CLIP、提示词结构和 VAE，这意味着整个流程无需反复加载或导入大型组件，就能实现无缝衔接，从而保持工作流程的轻量、高效和连贯。
缺点也很明显——ZIB 无法直接生成显式（XXX）内容。虽然在第二轮精炼时可以使用 Moody Porn Mix 等专用模型对私密部位进行重绘或增强，但这类场景下的姿势仍然不够稳定。不过，随着未来不久将推出更多专注于 XXX 内容的 ZIB LoRA，姿势问题应该很快得到解决。
这是一个实验性工作流程，很可能只是暂时的——直到官方的 ZIB 加速 LoRA 发布为止。但在目前阶段，这是一个实用且高效的方法，绝对值得一试。