ad
ad

ComfyUI better image upscaling

People & Blogs

Introduction

Hello everyone and welcome to a new video! Today's video is a continuation of the previous one about upscaling with ComfyUI images. In today's article, I’ll show you a handy trick to add more details to your image before you upscale it by 4X.

As shown below, you can see the difference in the workflow I’ll discuss. The standard upscaling we did yesterday is fast and efficient, but this new method is a bit slower and provides much more detailed images. You can observe enhanced details in the ship's decorations, the lighting, and the wave patterns. Overall, the image on the right showcases more refined and clearer details compared to the one on the left.

In this guide, I will walk you through the refined workflow:

Workflow Overview

I am using the Control Box, which is free to download on my Ko-fi page. The link is in the description. This tool helps organize all my workflow essentials and control everything from one place. For image generation, I am using the AGI Excel model, available for silver and gold members. It provides precise and high-quality images. In this example, I’ll create an image of a cat using Prompt Engineer as my AI expert.

Here’s the step-by-step process:

  1. Original Image Generation: Generate the initial image using the desired prompt.
  2. First Upscale Level - Image to Image: Input the image into the model to upscale by 1.5X using the image-to-image method. This directly upscales the original image with all its details and imperfections.
  3. Second Upscale Level - Latent Space Upscaling: Take the output of the first sampler and upscale it by 1.5X within the latent space instead of the image space. This approach ensures minimal differences but adds significant texture and detail.
    • Ensure the second sampler has a denoise value of 0.35. Adjust this as needed to control the difference in the images.
  4. Final Output - Ultra Sharp Upscale: Once the image is upscaled in the latent space, upscale it by 4X using the Ultra Sharp model.

The combination of 1.5X latent upscale plus the 4X upscale results in a beautifully detailed final image, reaching a resolution of 6,144 by 6,144 pixels.

Performance and Challenges

Using this method, the enhanced smaller details such as decorations, lighting, and reflections are significantly improved. However, one major drawback is that it is slower than the conventional method. I use an 8GB VRAM graphics card, which extends the generation time.

Example

Let's create a cat image to compare both outcomes:

Normal Image:

  • Generates an average image.

Enhanced Image:

  • Provides more depth in the lighting, better-detailed butterflies, clearer eye details in the cat, and highly visible reflections and foliage.

Additional Resources

  • For more detailed instructions on using the Ultra Sharp model and relevant tools, please refer to the previous video.
  • Visit my Ko-fi page for more goodies, and check out my newly launched NFT gallery linked in the video description.

I hope you find this method valuable for creating highly detailed images. It may be slower, but the quality payoff is more than worth it. If you have any questions, feel free to reach out on my Ko-fi page. Make sure to subscribe for future videos. Thank you for reading!

Keywords

  • ComfyUI
  • Image Upscaling
  • 4X Upscale
  • AGI Excel Model
  • Latent Space Upscaling
  • Image to Image
  • Ultra Sharp Upscale
  • Detail Enhancement

FAQ

Q1: What is the benefit of using the latent space for upscaling images?

  • A: Upscaling in the latent space preserves and enhances small details, providing a much sharper and more defined final image compared to traditional image space upscaling.

Q2: How does the Control Box help in the workflow?

  • A: Control Box helps organize and control all workflow elements from one place, making the entire process smooth and efficient.

Q3: Why is the suggested method slower?

  • A: The method is slower because it involves processing the image in multiple stages, especially during latent space upscaling, which requires more computational power and time.

Q4: What configuration is recommended for best results?

  • A: The denoise value should be set around 0.35 in the second sampler. An 8GB VRAM graphics card, or higher, is recommended for handling the upscaling without significant delays.

Q5: Where can I find the tools and models used in this method?

  • A: The Control Box tool is available for free on my Ko-fi page, while the AGI Excel model is available for silver and gold members. Detailed installation instructions are provided in the previous video linked in the article.