Want to Add More Detail? Try the Noise Method

Introduction

Hello, this is Easygoing.

I’ve always loved illustrations with lots of intricate detail. When I was in high school, I read the manga Ghost in the Shell and was blown away by the sheer amount of detail in every frame.

The artist must have spent days working on a single panel.

When I started generating AI images, I wanted to create the same kind of detailed illustrations. Since AI is a machine (although some might argue otherwise), it should excel at detailed work like this.

A Standard Illustration

Here’s a scene from the creation of Avatar Sirius.

First, let’s look at the illustration generated according to the initial prompt.

As always, this anime model delivers stunning results. The presence of the avatar is captured beautifully.

Now, let’s add more details.

Adding More Detail with the Prompt

First, I tried to increase the level of detail by adjusting the prompt.

I added highly detailed to the original prompt.

The details have increased significantly, but I wanted to push it further.

What is the Noise Method?

Now, let me explain the main focus of today’s post: the noise method.

You can read more about it from the creator’s detailed explanation at this link.

Since the page is in Japanese, please use your browser's translation function if necessary.

In simple terms, image generation AI restores images from noise. By continuously adding noise during the restoration process, you can generate an abundance of small objects, enhancing the level of detail.

The Effect of the Noise Method

Here’s what the noise method can do.

As you can see, the details have dramatically increased compared to the earlier version! The result feels more dynamic and I’m very satisfied.

How It Works

I followed the same method as the creator, but I’ll show you how to implement it using Stable Diffusion webUI Forge.

For this method, we’ll use a feature called ControlNet, which helps structure the composition of the illustration.

First, download the ControlNet lineart modelKatarag_lineartXL-fp16.safetensors from this link:

Place the downloaded Kataragi-LineartXL-fp16 file into your webuimodelsControlNet folder.

Next, download this noise image and save it somewhere on your computer.

Right-click (or long-press) on the image and select "Save Image As." Once everything is ready, launch Stable Diffusion webUI Forge.

1. First, open ControlNet.

2. Check ControlNet 0 and make sure it is enabled.
3. Upload the noise image.
4. Select Kataragi-LineartXL-fp16 as the model.
5. Set Control Weight to 0.5, Starting Control Step to 0.1, and Ending Control Step to 0.75.
6. For the resize option, choose Clip and Resize (this part is cut off in the image).

The other settings can be left as default. For a more detailed explanation of each setting, refer to the creator’s page linked earlier.

Once you have set everything up, you can save these settings as a preset and continue generating under the same conditions.

What Happens if You Increase the Noise?

If you raise the Control Weight beyond the previous setting, you can further increase the level of detail.

However, going too high can cause the composition to collapse.

In fact, the previous example was right on the edge of falling apart.

Let’s look at another example. Here, I’m using a prompt for a couple on a rooftop in a cityscape, which I used in the Refiner phase.

Beautiful sparkling lines have appeared, but this is actually a sign of collapse. If you keep increasing the strength…

The sparkles have taken over the composition! And there’s an extra person now!

AI often fills empty spaces with random elements—this is a common phenomenon in image generation.

The limit varies depending on the image and resolution, so feel free to experiment with your settings. Personally, I set mine to 0.7, which is just at the edge of collapse.

Combining with the Refiner

By using the noise method to increase detail and combining it with the Refiner to blur the background, I’ve been able to create illustrations that match my style.

As someone who couldn’t draw at all, this feels like I’ve become an illustrator exploring my own artistic style—it's exciting!

Thank you for reading all the way to the end!