AI Art Refinement: How Denoise Transforms Your Images

Introduction

Hello, this is Easygoing.

Today, we’ll take a look at image-to-image, a process for refining illustrations.

Text-to-Image vs. Image-to-Image

Image generation can be broadly divided into two types:

• Text-to-Image: Generating illustrations from text prompts.
• Image-to-Image: Generating illustrations from existing images.

Text-to-image involves inputting a description (prompt) to create an illustration.

In contrast, image-to-image involves inputting an image to refine an illustration, allowing various applications such as minor adjustments or significant redraws by tweaking the denoise strength.

Workflow Introduction!

This time, we’ll use a workflow combining both text-to-image and image-to-image.

In this workflow, we’ll create the base illustration with the versatile SDXL model and refine it with the high-texture Flux.1 model.

flowchart LR
subgraph SDXL
A1(Base Illustration<br>1024 x 1024)
end
B1(HDR Processing<br>SuperBreasts)
C1(upscale x1.4<br>Lanczos)
subgraph Flux.1
D1(Refined Illustration<br>1448 x 1448)
end
A1-->B1
B1-->C1
C1-->D1

• SDXL_Flux1_20250613.json

Models Used

• anima_pencil-XL-v5.0.0
• flux1-dev
• FLUX1-schnell-AE-FP32
• flan-t5-xxl_TE-only_FP32
• CLIP-SAE-ViT-L-14 -FP32

Lightweight Versions

• flux1-dev (GGUF)
• flan-t5-xxl_TE-only (FP16, FP8, GGUF)

Actual Illustrations

Let’s take a look at the actual illustrations.

SDXL (Base Illustration)

SDXL can produce clean anime illustrations with diverse compositions. However, compared to newer models, its noise removal efficiency is lower, resulting in slightly inferior texture and some residual noise in the final illustration.

Flux.1 (Refined Illustration)

Using Flux.1 for image-to-image processing on the SDXL base illustration, we achieved a higher overall texture quality and a clear illustration with minimal noise.

Denoise Adds Noise!

In image-to-image, the denoise parameter determines the strength of redrawing the image. While "denoise" translates to "noise removal" in Japanese, it actually controls the strength of adding noise.

Setting denoise to 0.5 means redrawing half of the illustration, but what exactly does this process entail?

Text-to-Image Case

In text-to-image, since the illustration is created from scratch, denoise is set to 1.0. With a karras scheduler and 10 steps, noise is reduced as follows:

The noise drops sharply at first and then gradually decreases in the later stages.

Image-to-Image Case

Now, let’s look at image-to-image with denoise set to 0.5. The result is shown in color, aligned to the right side of the previous graph.

The graph for denoise 0.5 resembles the latter half of the denoise 1.0 graph. Rather than uniformly reducing noise levels, denoise uses the latter part of the noise distribution.

Scheduler Differences Matter!

In the previous article, we compared scheduler differences with denoise set to 1.0:

Denoise 1.0, 10 Steps

With denoise at 1.0, all schedulers start from the same maximum noise level.

Denoise 0.5, 5 Steps

In contrast, with denoise set to 0.5, the initial noise amount varies significantly depending on the scheduler.

Changes in noise levels affect the strength of redrawing the illustration. In image-to-image, the scheduler differences have a much stronger impact compared to text-to-image.

Which Scheduler to Use?

So, which scheduler should you use for image-to-image?

In the previous article, we recommended two schedulers for text-to-image:

• Fast generation with SD1.5/SDXL → Karras
• Enhancing details with next-gen models → Beta

The same logic applies to image-to-image. For fast and stable generation, use the karras scheduler; for significant redrawing with next-gen models, use the beta scheduler.

Image-to-image is heavily influenced by scheduler differences, and model compatibility, so it’s also worth experimenting with different denoise and scheduler combinations to find the optimal setup as you gain experience.

Summary: Denoise Adds Noise!

• Denoise controls the strength of redrawing.
• Denoise adds noise.
• Denise levels significantly affects the outcome.

Image generation AI relies on illustrations from noise. Since its inception, many excellent methods have been developed to control noise.

Schedulers and denoise settings may seem complex, but understanding how they affect the outcome makes adjustments easier. The changes in illustrations when settings align perfectly are fascinating, and I look forward to exploring the best configurations while enjoying image generation.

Thank you for reading!