ComfyUI : Ultimate Upscaler - Upscale any image from Stable Diffusion, MidJourney, or photo!

Scott Detweiler
23 Aug 202309:02

TLDRIn this video, Scott Weller, the QA guy for stability.ai, demonstrates his technique for upscaling images using a calculator to determine the optimal size for SD Excel. He introduces custom nodes like the Control Net, Comfy Manager, Ultimate SD Upscaler, and a recommended resolution calculator to ensure the image matches SD Excel's trained resolutions. The process involves loading an image, using a checkpoint, and applying the upscaler with a denoising level set to 25%. The image is broken down into 1024x1024 tiles, and the upscale factor is calculated to upscale the image by four times. The video concludes with a successful upscale, showcasing the tool's ability to handle various image sources without the need for manual input of numbers or resolution adjustments.

Takeaways

  • 📈 Scott Weller, the QA guy for stability.ai, introduces an upscale technique for generating large images without manual input of numbers.
  • 🧩 He utilizes a calculator to determine the optimal size for SD Excel, which is trained in specific resolutions.
  • 📐 Custom nodes such as Control Net, Comfy Manager, Ultimate SD Upscaler, and a recommended resolution calculator are used in the process.
  • 🔍 The image is broken down into tiles, specifically 1024x1024 tiles, which is a key step in the upscaling process.
  • 📝 The use of a resolution calculator ensures that the upscaled image matches the recommended resolutions for SD Excel.
  • 🔢 A scale factor is used to determine the upscale size, which can be adjusted based on the desired outcome (e.g., 4 times upscale).
  • 🔄 The process involves multiplying the image's dimensions by the scale factor to get the target size for upscaling.
  • 🔍 The upscaler adds detail to the image, guided by the model and the amount of denoise applied.
  • 🛠️ The technique is applicable to any image from Stable Diffusion, MidJourney, or a photograph.
  • ⏱️ The upscaling process is time-consuming due to the large image size and the mathematical operations involved.
  • 💬 Weller encourages feedback in the comments and offers the graph to patrons at a higher support level on YouTube.

Q & A

  • What is the main purpose of the upscale technique discussed in the video?

    -The main purpose of the upscale technique is to increase the size of any image significantly without the need to manually input numbers or figure out specific settings, ensuring the image is upscaled to match the trained resolutions of models like SD Excel.

  • Who is the presenter of the video and what is his role?

    -The presenter of the video is Scott Weller, who is the QA guy for stability.ai.

  • What is the significance of using a calculator to determine the optimal size for upscaling?

    -Using a calculator ensures that the upscaled image matches the recommended resolutions that the model, SD Excel, has been trained on, which helps in achieving better upscaling results.

  • What are the custom nodes required for the upscaling process as mentioned in the video?

    -The custom nodes required are the Control Net, Comfy Manager, Ultimate SD Upscaler, Comfy Math, and a Recommended Resolution Calculator.

  • How does the Ultimate SD Upscaler break down the image for upscaling?

    -The Ultimate SD Upscaler breaks the image down into tiles, specifically into 1024 by 1024 pixel tiles, which is a key part of the upscaling process.

  • What is the role of the 'denoising' parameter in the upscaling process?

    -The denoising parameter controls how much of the upscaled image is allowed to change. A lower value like 0.25 means only 25% of the upscaled image is allowed to change, which helps in adding detail without introducing unwanted noise.

  • How does the upscaler handle the aspect ratio of the original image?

    -The upscaler uses the nearest SD Excel resolution from Comfy Math and then the Recommended Resolution Calculator to ensure the upscaled image maintains the correct aspect ratio and resolution.

  • What is the purpose of the 'reverse upscale' feature?

    -The 'reverse upscale' feature allows users to first upscale an image and then downscale it to a specific size. This can be useful if the desired final size is smaller than what the upscaler initially provides.

  • How does the presenter ensure that the upscaled image adds detail without being guided by a prompt?

    -The presenter uses a combination of the Ultimate SD Upscaler, the denoising parameter set to a low value, and the model's inherent understanding guided by the prompts and the amount of denoise to add detail to the upscaled image.

  • What is the expected time for the upscaling process to complete?

    -The upscaling process is not quick and can take a significant amount of time, as demonstrated in the video where it took 354 seconds to complete.

  • How can viewers access the graph used in the video for their own use?

    -Viewers who are at the sponsor level or higher on the YouTube channel can access the graph in the private area. Others can build the graph themselves by following the instructions in the video.

  • What type of images can be upscaled using the technique shown in the video?

    -The technique can upscale any type of image, whether it's from Stable Diffusion, MidJourney, or a photograph.

Outlines

00:00

📈 Upscaling Images with SD Excel: A Tutorial

Scott Weller, the QA guy for stability.ai, introduces his technique for upscaling images to larger sizes using a calculator to determine the optimal size for SD Excel, which is trained in certain resolutions. He explains the process of using the Ultimate SD Upscaler tool, which breaks down images into tiles, and the importance of matching SD Excel's expected tile size of 1024x1024. Scott also details the use of a recommended resolution calculator to ensure the upscaled image matches the resolutions that SD Excel works best with. He guides viewers through setting up the process, including using custom nodes like Control Net and Comfy Manager, and emphasizes the ease of using the system without needing to manually input numbers.

05:01

🔍 Advanced Image Upscaling and Resolution Matching

The second paragraph delves deeper into the image upscaling process, emphasizing the use of an upscale factor to determine the desired resolution for the upscaled image. Scott discusses the use of a reverse upscale feature, which allows for fine-tuning the upscaling process. He also explains the rationale behind using the 'nearest sdxl resolution' from the Comfy Math node and then putting it into the recommended resolution calculator, highlighting the flexibility of the process to accommodate different scaling factors. The summary concludes with Scott's satisfaction with the upscaling results and his invitation for viewers to share their thoughts in the comments section. He also mentions the availability of the graph for sponsors and encourages viewers to build and save the graph for their use.

Mindmap

Keywords

💡Upscale

Upscaling refers to the process of increasing the resolution of an image or video. In the context of the video, it involves enhancing the quality of an image to a higher resolution without losing detail or clarity. The video demonstrates a technique for upscaling images using specific software tools to achieve larger, more detailed images.

💡Stable Diffusion (SD)

Stable Diffusion is a term that likely refers to a specific AI model or software used for image processing, in this case, for upscaling images. The video mentions matching resolutions that SD is trained in, indicating that the process is optimized for the capabilities of this particular model.

💡Tiles

In the context of image processing, tiles refer to the segments or portions of an image that are processed individually. The video discusses breaking down an image into tiles of 1024 by 1024 pixels, which is a method used to manage and process large images more efficiently.

💡ComfyUI

ComfyUI seems to be a user interface or a tool mentioned in the video that is used for managing and customizing the upscaling process. It is presented as a convenient way to handle image upscaling without needing to manually input complex data.

💡Ultimate SD Upscaler

The Ultimate SD Upscaler is a tool or feature within the software being discussed that is used to upscale images. It is described as a fantastic tool that breaks down images into tiles and enhances them, suggesting it is a key component of the upscaling technique shown.

💡Resolution Calculator

A resolution calculator is a tool used to determine the optimal size for images to match specific software's recommended resolutions. In the video, it is emphasized as a crucial part of the upscaling process to ensure compatibility and quality of the upscaled images.

💡Denoising

Denoising is a process that reduces or eliminates noise in images, which can be visual artifacts that detract from the quality. The video mentions a denoising level of 25 percent, which means that only a quarter of the upscaled image's changes are allowed to add detail, maintaining the original image's integrity.

💡Checkpoint

In the context of this video, a checkpoint likely refers to a saved state or specific configuration within the image processing software that the user can load to start the upscaling process. It is part of the workflow to ensure consistency and repeatability of the upscaling technique.

💡Custom Nodes

Custom nodes are user-defined functions or modules within a software that can be used to perform specific tasks. The video script mentions using custom nodes for control and management of the upscaling process, indicating a level of customization and flexibility in the software.

💡Image Processing

Image processing involves the manipulation of images using various algorithms and techniques to achieve desired effects. The video is centered around image processing, specifically upscaling images to higher resolutions with the aid of AI and specific software tools.

💡Mathematical Operations

Mathematical operations are calculations performed on data, and in the context of this video, they are essential for determining the correct scaling and resolution adjustments needed for upscaling images. The use of 'Comfy Math' suggests a reliance on mathematical precision to achieve high-quality results.

Highlights

Scott Weller demonstrates an upscale technique for creating large images using SD Excel.

The method involves using a calculator to determine the optimal size for images to match SD Excel's training resolutions.

Custom nodes such as Control Net, Comfy Manager, Ultimate SD Upscaler, and Recommended Resolution Calculator are utilized.

The Ultimate SD Upscaler breaks down images into tiles for processing, using a preferred tile size of 1024x1024.

The process includes creating positive and negative blanks for guiding the upscale and adding detail.

A denoising level of 25% is suggested to allow for detail addition without excessive random changes.

The Recommended Resolution Calculator is key for ensuring images match SD Excel's recommended resolutions.

The image's width and height are multiplied by a scale factor, typically set to 4 for a four times upscale.

An upscale model is used to add skin texture and other details to the image.

The process is automated to avoid manual input of numbers, simplifying the upscale for any image.

The upscale factor can be adjusted for different scaling needs, such as 1.5 or 1.7 times.

The process involves breaking down the image into 1024x1024 tiles, which is a crucial part of the upscaling secret.

The upscaled image is unguided by prompts, relying on the model and denoise levels for detail.

The process is time-consuming but results in a high-quality upscale, with an example taking 354 seconds.

The technique is applicable to any image type, including Stable Diffusion, MidJourney, or photographs.

The upscaled image is created in 16 tiles, depending on the input image's size.

The process ensures that the proper resolution for SD Excel is used for optimal results without manual calculations.

The graph used in the demonstration will be available for sponsors to download from a private area on YouTube.