Sampling method - Euler A Substep

什么是 Euler A Substep？AI 绘画背后的“导航员”

What is Euler A Substep? The “Navigator” Behind AI Art

在探索 AI 绘画（特别是像 Stable Diffusion 这样的扩散模型）的世界时，你可能会遇到各种各样令人眼花缭乱的参数。其中，“采样器（Sampler）” 是最重要的设置之一，而 Euler A 则是最受用户喜爱的一种。

最近，随着技术的进步，Euler A Substep 作为一个更精细的概念出现了。今天，我们就用最通俗易懂的方式，为您揭开它的神秘面纱。

When exploring the world of AI art generation (especially distinct models like Stable Diffusion), you might encounter a dazzling array of parameters. Among them, the “Sampler” is one of the most critical settings, and Euler A is a fan favorite.

Recently, with technological advancements, Euler A Substep has emerged as a more refined concept. Today, let’s unveil its mystery in the most accessible way possible.

第一部分：基础概念——什么是“去噪”？

Part 1: The Basics — What is “Denoising”?

要理解 Euler A Substep，首先要理解 AI 是如何画画的。目前的 AI 绘画主要使用“扩散模型（Diffusion Model）”。

想象一下，你有一张清晰的照片，然后你往上面撒了一层又一层的沙子（这就是“噪声”），直到最后这张照片完全由沙子覆盖，变成了一片混沌的电视雪花点。

AI 的工作就是逆转这个过程： 它从一堆随机的沙子（噪声）开始，一点点把沙子扫掉，试图还原出原本并不存在、但它“脑补”出来的美丽画作。这个扫沙子的过程，就叫去噪（Denoising），也就是我们说的采样（Sampling）。

To understand Euler A Substep, we first need to understand how AI paints. Current AI art mostly uses “Diffusion Models.”

Imagine you have a clear photograph, and you sprinkle layer after layer of sand (this is the “noise”) onto it until the photo is completely covered, becoming a chaotic field of TV static.

The AI’s job is to reverse this process: It starts with a pile of random sand (noise) and sweeps it away bit by bit, attempting to reveal a beautiful image that didn’t exist before but is “imagined” by the AI. This process of sweeping away the sand is called Denoising, which we technically refer to as Sampling.

第二部分：谁在指挥扫沙子？——采样器（Sampler）

Part 2: Who Directs the Sweeping? — The Sampler

既然要把沙子扫掉，就需要一个“指挥官”告诉 AI 每一步该怎么扫、往哪个方向扫、扫多少。这个指挥官就是采样器（Sampler）。

  • Euler：这是一位非常传统的数学家指挥官。他做事直来直去，看准一个方向就走一步，简单高效。
  • Euler A (Ancestral)：这是 Euler 的一个更灵活的堂兄弟。他在每走一步之后，会稍微加入一点点“随机性（Randomness）”。这让生成的图片更有创意，不会每次都一模一样，但有时画面会因为这种随机性而处于一种“未完成”或“朦胧”的状态。

Since the sand needs to be swept away, we need a “Commander” to tell the AI how to sweep at each step, in which direction, and how much. This commander is the Sampler.

  • Euler: This is a very traditional mathematician commander. He works in a straightforward manner; he calculates a direction and takes a step. Simple and efficient.
  • Euler A (Ancestral): This is Euler’s more flexible cousin. After taking each step, he adds a tiny bit of “randomness.” This makes the generated images more creative and ensures they aren’t identical every time, but sometimes the image can feel “unfinished” or “hazy” due to this shifting nature.

第三部分：核心主角——什么是 Substep？

Part 3: The Core Concept — What is a Substep?

现在我们进入正题。如果你是个完美主义者，觉得 Euler A 虽然快，但有时候画得不够细腻，边缘不够清晰，该怎么办？

这就涉及到了 Substep（子步数/细分步数）。

形象的比喻：走迷宫

Metaphor: Navigating a Maze

想象 AI 生成图片就像是在迷宫中从起点（满是噪声）走到终点（清晰图片）。如果我们要走 20 步到达终点：

  1. 标准的 Euler A：
    指挥官每走一大步，就会重新观察一次地图，然后决定下一步怎么走。这很大大咧咧，走得快，但如果每一步跨度太大（特别是步数设定较低时），可能会走偏，导致画面细节丢失或结构错乱。

  2. 引入 Substep (细分步数)：
    Substep 就像是在原来的每一大步中间，增加了微小的调整步伐。
    指挥官说：“这 20 大步不变，但在迈出每一大步的过程中，我要在那一大步里再细分为几个小碎步来确认路线。”

    • 没有 Substep： 看地图 -> 迈一大步 -> 看地图 -> 迈一大步。
    • 有 Substep： 看地图 -> 迈一小步 -> 修正方向 -> 再迈一小步 -> 修正方向 -> （完成原来的一大步）。

Euler A Substep 就是在告诉 AI：在使用 Euler A 这个富有创意的算法时，请你在每一步计算中更加谨慎，进行额外的数学修正（通常涉及微分方程求解时的更高阶近似），从而让那“充满随机性”的一步走得更稳、更准。

Now to the main topic. If you are a perfectionist and feel that while Euler A is fast, it sometimes paints without enough subtlety or clear edges, what can you do?

This brings us to Substep.

Metaphor: Navigating a Maze

Imagine AI generating an image is like walking through a maze from the Start (Full Noise) to the Finish (Clear Image). If we plan to take 20 steps to reach the end:

  1. Standard Euler A:
    The commander makes a large stride, then re-checks the map to decide the next move. This is bold and fast, but if the stride is too wide (especially with low step counts), it might veer off course, leading to lost details or structural errors.

  2. Introducing Substep:
    Substep is like adding tiny adjustment steps within each original large stride.
    The commander says, “The 20 main steps remain, but while taking each large stride, I want to divide it into several tiny shuffle steps to confirm the route.”

    • No Substep: Check map -> Take big stride -> Check map -> Take big stride.
    • With Substep: Check map -> Take small step -> Correct course -> Take another small step -> Correct course -> (Complete the original big stride).

Euler A Substep essentially tells the AI: When using the creative Euler A algorithm, please be more cautious in each calculation step. Perform extra mathematical corrections (usually involving higher-order approximations in solving differential equations) so that the “random-filled” step is taken more steadily and accurately.

第四部分：为什么我们需要它？

Part 4: Why Do We Need It?

在 AI 技术社区（如 Stable Diffusion WebUI 或 ComfyUI 的讨论中），引入 Substep 主要是为了解决两个问题：

  1. 提升低步数下的质量： 以前你可能需要跑 30-50 步才能得到一张好图。有了 Substep 的辅助，可能只需要跑 10-15 步，画面就已经非常结实、清晰了，因为它每一步的质量都变高了。
  2. 解决“闪烁”与不稳定性： Euler A 有个特点是画面变化大。Substep 能够约束这种变化，让画面既保留 Euler A 的柔和感，又拥有类似 Euler 或 Heun 那样的准确度。

简单总结：

  • Euler A： 快速、有创意、偶尔随性、画面柔和。
  • Euler A + Substep： 在保留创意的基础上，变得更严谨、细节更丰富、线条更扎实。就像是给那位随性的画师配了一副高精度的眼镜。

In the AI community (such as discussions around Stable Diffusion WebUI or ComfyUI), introducing Substep is mainly to solve two problems:

  1. Improving Quality at Low Steps: Previously, you might have needed 30-50 steps to get a good image. With the assistance of Substep, you might only need 10-15 steps to get an image that is solid and clear, because the quality of each step has increased.
  2. Fixing “Flickering” and Instability: A characteristic of Euler A is significant variance. Substep constrains this variance, allowing the image to retain Euler A’s softness while possessing the accuracy of samplers like Euler or Heun.

Summary:

  • Euler A: Fast, creative, occasionally spontaneous, soft visuals.
  • Euler A + Substep: Creative but rigorous, richer in detail, and with more solid lines. It’s like giving that spontaneous artist a pair of high-precision glasses.

结论

Conclusion

Euler A Substep 并不是一种全新的魔法，它是对现有工具的一种精细化打磨。

如果把生成图片比作磨刀：

  • 步数 (Steps) 是你在磨刀石上磨多少次。
  • Euler A 是磨刀的手法。
  • Substep 就是在每次推刀的时候，更加细致地控制力度和角度，而不只是一股脑推过去。

理解了这一点，下次你在 AI 软件中看到这个选项时，不妨试着调高一点点，或许你会惊讶地发现，原本模糊的背景变得清晰了，人物的眼神也变得更加深邃了。

Euler A Substep is not a brand-new magic; it is a refined polish of existing tools.

If generating an image is like sharpening a knife:

  • Steps are how many times you glide the knife over the whetstone.
  • Euler A is the technique used to move your hand.
  • Substep is controlling the pressure and angle meticulously during each glide, rather than just shoving it forward blindly.

Understanding this, the next time you see this option in your AI software, try tweaking it up a bit. You might be surprised to find that a previously blurry background becomes sharp, and a character’s gaze becomes deeper and more profound.

What is Euler A Substep? The “Navigator” Behind AI Art

When exploring the world of AI art generation (especially distinct models like Stable Diffusion), you might encounter a dazzling array of parameters. Among them, the “Sampler” is one of the most critical settings, and Euler A is a fan favorite.

Recently, with technological advancements, Euler A Substep has emerged as a more refined concept. Today, let’s unveil its mystery in the most accessible way possible.

Part 1: The Basics — What is “Denoising”?

To understand Euler A Substep, we first need to understand how AI paints. Current AI art mostly uses “Diffusion Models.”

Imagine you have a clear photograph, and you sprinkle layer after layer of sand (this is the “noise”) onto it until the photo is completely covered, becoming a chaotic field of TV static.

The AI’s job is to reverse this process: It starts with a pile of random sand (noise) and sweeps it away bit by bit, attempting to reveal a beautiful image that didn’t exist before but is “imagined” by the AI. This process of sweeping away the sand is called Denoising, which we technically refer to as Sampling.

Part 2: Who Directs the Sweeping? — The Sampler

Since the sand needs to be swept away, we need a “Commander” to tell the AI how to sweep at each step, in which direction, and how much. This commander is the Sampler.

  • Euler: This is a very traditional mathematician commander. He works in a straightforward manner; he calculates a direction and takes a step. Simple and efficient.
  • Euler A (Ancestral): This is Euler’s more flexible cousin. After taking each step, he adds a tiny bit of “randomness.” This makes the generated images more creative and ensures they aren’t identical every time, but sometimes the image can feel “unfinished” or “hazy” due to this shifting nature.

Part 3: The Core Concept — What is a Substep?

Now to the main topic. If you are a perfectionist and feel that while Euler A is fast, it sometimes paints without enough subtlety or clear edges, what can you do?

This brings us to Substep.

Metaphor: Navigating a Maze

Imagine AI generating an image is like walking through a maze from the Start (Full Noise) to the Finish (Clear Image). If we plan to take 20 steps to reach the end:

  1. Standard Euler A:
    The commander makes a large stride, then re-checks the map to decide the next move. This is bold and fast, but if the stride is too wide (especially with low step counts), it might veer off course, leading to lost details or structural errors.

  2. Introducing Substep:
    Substep is like adding tiny adjustment steps within each original large stride.
    The commander says, “The 20 main steps remain, but while taking each large stride, I want to divide it into several tiny shuffle steps to confirm the route.”

    • No Substep: Check map -> Take big stride -> Check map -> Take big stride.
    • With Substep: Check map -> Take small step -> Correct course -> Take another small step -> Correct course -> (Complete the original big stride).

Euler A Substep essentially tells the AI: When using the creative Euler A algorithm, please be more cautious in each calculation step. Perform extra mathematical corrections (usually involving higher-order approximations in solving differential equations) so that the “random-filled” step is taken more steadily and accurately.

Part 4: Why Do We Need It?

In the AI community (such as discussions around Stable Diffusion WebUI or ComfyUI), introducing Substep is mainly to solve two problems:

  1. Improving Quality at Low Steps: Previously, you might have needed 30-50 steps to get a good image. With the assistance of Substep, you might only need 10-15 steps to get an image that is solid and clear, because the quality of each step has increased.
  2. Fixing “Flickering” and Instability: A characteristic of Euler A is significant variance. Substep constrains this variance, allowing the image to retain Euler A’s softness while possessing the accuracy of samplers like Euler or Heun.

Summary:

  • Euler A: Fast, creative, occasionally spontaneous, soft visuals.
  • Euler A + Substep: Creative but rigorous, richer in detail, and with more solid lines. It’s like giving that spontaneous artist a pair of high-precision glasses.

Conclusion

Euler A Substep is not a brand-new magic; it is a refined polish of existing tools.

If generating an image is like sharpening a knife:

  • Steps are how many times you glide the knife over the whetstone.
  • Euler A is the technique used to move your hand.
  • Substep is controlling the pressure and angle meticulously during each glide, rather than just shoving it forward blindly.

Understanding this, the next time you see this option in your AI software, try tweaking it up a bit. You might be surprised to find that a previously blurry background becomes sharp, and a character’s gaze becomes deeper and more profound.