ULTIMATE Upscale for SLOW GPUs - Fast Workflow, High Quality, A1111

Olivio Sarikas
6 May 202313:48

TLDRIn this video, the host introduces a solution for slow GPUs to achieve high-quality image upscaling using an extension called 'Ultimate Upscale' for the software Automatic 1111. The process involves splitting the image into 512x512 tiles, applying settings, and then recombining them for a superior result. The video provides a step-by-step guide on installing the extension, using the 4X Ultra Sharp Upscaler model, and adjusting settings for optimal results. The host also demonstrates the difference between standard upscaling and the Ultimate Upscale method, showing significant improvements in detail and resolution. Additionally, the video covers techniques for refining the upscaling process, such as using ControlNet 1.1 for more detailed results and inpainting to fix details like eyes. The summary emphasizes the balance between denoise strength for detail preservation and the potential for changes in the original image. The host concludes with a teaser for a live stream and encourages viewers to like the video.

Takeaways

  • 📈 Use the 'Ultimate Upscale' extension for Automatic 1111 to upscale images from a low to a higher resolution with better quality.
  • 🔗 Install the extension by copying the provided web address and using the 'Install from URL' option in the Extensions tab of Automatic 1111.
  • 📚 After installation, always check for updates and restart the UI to ensure the new model is loaded.
  • 📈 For even higher quality, consider using the 'Ultra Sharp Upscaler' method, which is detailed in another video by the same creator.
  • 📂 Download and install the 4X Ultra Sharp Upscaler model into the appropriate folder within your Automatic 1111 directory.
  • 📌 When using the Ultimate SD Upscale in the Image-to-Image tab, ensure to use the same model that was used to create the original image.
  • ⚙️ Adjust the settings by removing specific image details from the positive prompt and turning off 'Restore Face' for tiled images.
  • 🔍 Experiment with different sampling methods like Euler a, DPM2 Keras, or DPM++ 2sa Keras to find the best results.
  • 🔍 Set the denoise strength between 0.2 and 0.3 for most images, but be aware that higher values can alter the original image.
  • 🖼️ For anime-style images, use the rESRGAN 4X Anime model for upscaling to maintain the style's characteristics.
  • 👁️‍🗨️ If the upscaled image has issues with facial details, use the 'Inpaint' feature to manually enhance those areas.
  • ⏱️ Be prepared for longer render times when using ControlNet with tiling and the Ultimate Upscale extension, as they provide more detail but take more time.

Q & A

  • What is the main issue with using the 'extras' tab in automatic 1111 for upscaling?

    -The main issue is that it provides an acceptable quality when upscaling from a small resolution image to a higher resolution image, but the quality is not great.

  • What is the recommended extension for automatic 1111 to improve upscaling quality?

    -The recommended extension is called 'ultimate upscale' which splits the image into tiles of 512 by 512 and applies settings to them for higher quality.

  • How can one install the ultimate upscale extension in automatic 1111?

    -You go to the extensions tab, select 'install from URL', paste the provided web address, and then click on the install button.

  • What is the suggested upscaling method for images that require high quality?

    -The suggested method is to use the 4X Ultra sharp upscaler, which is available for download and provides a high-quality upscaling.

  • What is the purpose of using the 'ultimate SD upscale' script in the image to image tab?

    -The script is used to upscale images with additional options, providing a more detailed and higher quality output compared to standard upscaling methods.

  • Why is it important to check the model used to create the image before upscaling?

    -It's important to use the same model for upscaling to ensure consistency and compatibility, avoiding potential issues with the final output quality.

  • What adjustments are necessary when using the ultimate upscale extension?

    -You need to remove specific information from the positive prompt field, turn off the 'restore face' option, and adjust the denoise strength and sampling method for optimal results.

  • What is the recommended denoise strength for most purposes when using the ultimate upscale extension?

    -The recommended denoise strength is between 0.2 and 0.3, which provides a balance between detail and maintaining the original image characteristics.

  • How does the final output differ when using the 'r ESR gen 4X anime' model for upscaling?

    -The 'r ESR gen 4X anime' model is designed for anime-style images and drawings, providing more detail and a higher resolution output that is more suitable for these types of images.

  • What is the process for fixing blurry or incorrectly rendered eyes in an upscaled image?

    -After the upscale render, the image is sent to 'in paint' where the eyes are masked out and then inpainted with a 'beautiful eyes' prompt to achieve a more realistic and detailed result.

  • Why is the render time longer when using control net with the ultimate upscale extension?

    -Using control net with the ultimate upscale extension involves processing each tile individually, which can significantly increase the render time, especially for images with high resolution or complex details.

Outlines

00:00

🖼️ Enhancing Image Quality with Ultimate Upscale

The first paragraph introduces the viewer to a method for improving the quality of upscaled images, particularly for those with slower GPUs. The host recommends using the 'ultimate upscale' extension for Automatic 1111, which divides images into 512x512 tiles and applies settings to enhance quality. The process involves installing the extension from a provided URL, updating, and restarting the application. The host also suggests using the 'ultra sharp upscaler' for even better results. Detailed instructions are given for installing and using the upscaler, including downloading models and adjusting settings such as denoise strength and sampling methods.

05:02

🎨 Balancing Detail and Originality in Upscaling

The second paragraph delves into the nuances of using the upscaler, discussing the trade-offs between detail and maintaining the original image's integrity. It explains how setting the denoise strength can affect the final image's appearance, with higher values potentially altering the original. The paragraph also covers different methods for using the upscaler, such as using ControlNet 1.1 for more detailed results, albeit with a longer render time. Examples are provided to illustrate the differences in image quality and detail achieved through various settings and methods.

10:03

👀 Fixing Faces in Upscaled Images

The third paragraph addresses a common issue with upscaling faces, where the eyes may appear blurry or distorted. The host demonstrates a technique to correct this by using a masking tool to isolate the eyes and then using the 'in paint' feature to enhance the eye details. The paragraph also explores the use of ControlNet with tile processing for a more detailed result, albeit with a longer render time. The host emphasizes the importance of finding a balance between detail and the original image's appearance, and concludes with a call to action for viewers to like the video and join a live stream.

Mindmap

Keywords

💡Upscaling

Upscaling refers to the process of increasing the resolution of an image or video. In the context of the video, upscaling is used to enhance the quality of images from a lower to a higher resolution, which is particularly useful for images that are intended to be displayed on larger screens or for those requiring more detail.

💡Slow GPUs

Slow GPUs, or Graphics Processing Units, are components in computers that are responsible for rendering images, animations, and videos. The term 'slow' implies that these GPUs may not be as efficient or fast as their more powerful counterparts. The video discusses techniques to improve image quality even when using less powerful hardware.

💡Ultimate Upscale Extension

The Ultimate Upscale Extension is a tool mentioned in the video that is used to improve the quality of upscaled images. It works by dividing an image into tiles and then applying specific settings to each tile to achieve a higher quality result. This is significant as it allows users to achieve better image quality despite having slow GPUs.

💡Tiles

In the context of the video, tiles refer to the segments into which an image is divided for the purpose of upscaling. The Ultimate Upscale Extension splits an image into 512 by 512 pixel tiles. This method is beneficial as it allows for more detailed processing and can result in a higher quality final image.

💡Automatic 1111

Automatic 1111 appears to be a software or platform mentioned in the video where the upscaling process is carried out. It is where the Ultimate Upscale Extension is installed and used to enhance image quality. The script and settings discussed in the video are all applied within this software.

💡Denoise Strength

Denoise Strength is a parameter that controls the level of noise reduction in an image. A higher denoise strength can result in more detail and better quality, but it may also alter the original image. In the video, it is suggested to use a denoise strength between 0.2 and 0.3 for most purposes.

💡ControlNet 1.1

ControlNet 1.1 is a model or tool used for image processing, specifically mentioned for its ability to add more detail to upscaled images. It is used in conjunction with the Ultimate Upscale Extension to process each tile of the image individually, which can enhance the detail but also requires more processing power and time.

💡ESRGAN

ESRGAN, which stands for Enhanced Super-Resolution Generative Adversarial Network, is a type of model used for upscaling images. In the video, it is used to upscale images to 4 times their original size with a focus on quality, particularly for photographs to add details to the skin and hair.

💡Image-to-Image

Image-to-Image is a process mentioned in the video where an initial image is transformed into another through a series of steps. It involves using the Ultimate Upscale Extension and may also involve using ControlNet 1.1 for additional detail. It is part of the workflow for achieving high-quality upscaled images.

💡Restore Face

Restore Face is a feature within the upscaling process that attempts to improve the clarity and detail of faces in images. However, the video suggests turning this feature off when tiling images as it may fail due to the full face not being visible in each tile. The feature is re-enabled later for specific parts of the image, like the eyes, to fix any blurriness.

💡Inpainting

Inpainting is a technique used to fill in or repair parts of an image. In the video, it is used to improve the quality of the eyes in an image after upscaling, where the eyes may appear blurry. The process involves masking out the area to be improved and then generating new details to replace the original, resulting in a more refined image.

Highlights

The video introduces a method to upscale images for slow GPUs using an extension called Ultimate Upscale for Automatic 1111.

Ultimate Upscale splits images into 512x512 tiles and applies settings to enhance quality.

The process involves installing the Ultimate Upscale extension from a provided URL in Automatic 1111.

Ultra Sharp Upscaler is recommended for an additional upscaling method.

The 4X Ultra Sharp Upscaler model is used for high-quality upscaling.

The model database in Automatic 1111 is used to select the appropriate model for upscaling.

Adjustments are made to the positive and negative prompts when using tiling for upscaling.

The restore face feature is turned off due to tiling, as it may fail to process the full face.

Euler A with 20-30 steps is suggested for sampling methods.

Denoise strength between 0.2 and 0.3 is recommended for most cases.

ControlNet 1.1 can be used for more detailed upscaling, but may require more time.

Examples are provided to compare the quality of different upscaling methods.

Anime style images can be upscaled using the RESRGen 4X Anime model.

Face details can be improved by masking and inpainting with specific prompts.

ControlNet with tile and high denoise strength can yield intricate details in images.

The video concludes with a live stream announcement and a call to action for likes.