English

This is a workflow designed to generate 4K images through a dual-pass sampling process. The recommended resolution is 1080×1920 or 1920×1080, but you can adjust it slightly higher or lower depending on your needs.

In the first pass, generate a satisfying base image. Then, use the second pass for 2× upscaling to achieve highly detailed 4K images with much clearer textures. Compared with Seed VR2, this method can produce significantly sharper and more refined details.

However, I should point out that this workflow is much more time-consuming than using Seed VR2 for upscaling. It is essentially a trade-off: more processing time in exchange for higher image quality.

The UNet can be replaced with any version of Krea2 Turbo.

中文

这是一个通过双采样流程生成 4K 图像的工作流。推荐初始分辨率设置为 1080×1920 或 1920×1080，也可以根据需求适当调整大小。

在第一次采样中，先生成一张满意的基础图像；随后通过第二次采样进行 2 倍放大与细节重构，最终可以获得纹理非常精细的 4K 图像。相比 Seed VR2，这种方法能够保留更加清晰的纹理细节，尤其是在放大查看原图时，差异会更加明显。

不过需要说明的是，这种方式相比使用 Seed VR2 进行放大，会消耗更多时间。它本质上是用更长的生成时间换取更高的图像质量，适合追求最终细节表现的场景。

UNet 可以替换为任意版本的 Krea2 Turbo 模型。