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

Olivio Sarikas
6 May 202313:48

TLDRIn this video, the host introduces a method to upscale images for users with slow GPUs, using an extension called 'Ultimate Upscale' for the software 'Automatic 1111'. The extension divides images into 512x512 tiles, applying settings to enhance quality before recombining them. The process involves installing the extension from a provided URL, updating, and restarting the UI. The host also recommends the 'Ultra Sharp Upscaler' for even higher quality and guides viewers on how to download and install it. Different models like Realistic Vision version 2 and RESRGen 4X are discussed, along with settings adjustments for optimal results. Examples are shown comparing the original image to the upscaled versions using different methods and settings. The video concludes with a live stream announcement and an invitation for viewers to engage with the content.

Takeaways

  • 🔍 **Using Ultimate Upscale for Automatic 1111**: To improve image quality on slow GPUs, use the Ultimate Upscale extension which splits images into 512x512 tiles for higher quality upscaling.
  • 🌐 **Installation Process**: Install the Ultimate Upscale extension from the provided URL in the Automatic 1111 extensions tab.
  • 📚 **Additional Upscaling Method**: Consider using the Ultra Sharp Upscaler for another high-quality upscaling method, as demonstrated in a separate video.
  • 📁 **Model Database**: Download and install the 4X Ultra Sharp Upscaler model into the Automatic 1111's models folder for enhanced upscaling.
  • 📈 **Image to Image Tab**: For upscaling, navigate to the 'Image to Image' tab and select 'Ultimate SD Upscale' from the script options.
  • 🔄 **Tiling Adjustments**: When upscaling, adjust the positive prompt to neutral information and turn off 'Restore Face' due to the tiling process.
  • ⚙️ **Sampling Methods**: Experiment with different sampling methods like Euler a, DPM2 Keras, or DPM++ 2sa Keras to find the best results.
  • 🔍 **Denoise Strength**: Set the denoise strength between 0.2 and 0.3 for most images to balance detail and original image integrity.
  • 🖼️ **Upscaling Anime Style Images**: For anime-style images, use the RESRGen 4X Anime model for more detailed upscaling.
  • 🎨 **Post-Upscaling Touch-ups**: After upscaling, use tools like 'Inpaint' to fix details like eyes that may not have rendered correctly.
  • ⏱️ **Render Time Consideration**: Using ControlNet with tiling and the Ultimate Upscale extension can significantly increase render time but may yield better results.

Q & A

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

    -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 to improve upscaling quality in automatic 1111?

    -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 you install the ultimate upscale extension in automatic 1111?

    -You can install it by going to the extensions tab, selecting 'install from URL', pasting the provided web address, and clicking on the install button.

  • What is the role of the '4X Ultra sharp upscaler' in the upscaling process?

    -The '4X Ultra sharp upscaler' is another upscaling method that provides high quality results and is suggested for use in conjunction with the ultimate upscale extension.

  • How does the ultimate upscale extension work with the image?

    -It works by splitting the image into tiles, applying settings to each tile individually, and then recombining the tiles to form the final high-quality image.

  • What should be considered when adjusting the positive and negative prompts for the ultimate upscale script?

    -You should remove specific information related to the original image to avoid conflicts with the tiling process. For example, if the original prompt included 'highly detailed skin', this should be removed.

  • Why is the 'restore face' option turned off during the tiling process?

    -The 'restore face' option is turned off because the full face is not visible in any single tile during the tiling process, and it is likely to fail or produce undesirable results.

  • What is a good range for the Euler a sampling method steps?

    -The video suggests that an Euler a sampling method with between 20 and 30 steps is good, with 25 steps being suitable for most purposes.

  • How does the denoise strength setting affect the final image quality?

    -A higher denoise strength can produce more details and better quality, but it may also alter the image or cause parts of it to not fit the original. A balance needs to be found, with the video suggesting a denoise strength between 0.2 and 0.3.

  • What is the purpose of using ControlNet 1.1 in conjunction with the ultimate upscale script?

    -ControlNet 1.1 is used to add more detail to each of the 512 by 512 tiles individually, which can result in a more detailed final image, although it may take longer to render.

  • How can one fix issues with facial features, such as blurry eyes, after using the ultimate upscale script?

    -After the upscale render, the image can be sent to 'In Paint' to manually mask out the eyes and repaint them for better detail and clarity.

Outlines

00:00

🖼️ Upscaling Techniques for Low-End GPUs

The video introduces viewers to methods for enhancing image quality using automatic 1111's 'ultimate upscale' extension, which divides images into 512x512 tiles for better upscaling. The host also recommends the 'ultra sharp upscaler' for even higher quality and provides a step-by-step guide on installing and using these tools, including downloading models and adjusting settings for optimal results.

05:02

🔍 Enhancing Image Details with Upscaling Scripts

The second paragraph delves into the specifics of using the 'ultimate upscale' script, discussing the importance of selecting the correct model and adjusting settings such as denoise strength and sampling methods. The host compares different upscaling results, including those from the 'ultra sharp' model and 'image to image' match, and explains how to avoid unwanted changes in the image through careful prompt adjustments.

10:03

👁️‍🗨️ Fixing Faces and Eyes in Upscaled Images

The final paragraph addresses the common issue of faces appearing blurry or distorted after upscaling. The host suggests a solution involving in-painting the eyes after the initial render to achieve a more natural look. Additionally, the use of 'control net 1.1' for even more detailed results is mentioned, with an emphasis on the trade-off between render time and image quality. The video concludes with a call to action for viewers to like the video and join the upcoming 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 a low-resolution image to a higher resolution, which is crucial for achieving better visual results. The video discusses using an extension called 'ultimate upscale' to achieve high-quality upscaling, particularly beneficial for users with slow GPUs.

💡Slow GPUs

Slow GPUs are graphics processing units that do not perform at high speeds, which can limit the capabilities of rendering and upscaling images. The video is aimed at users with such hardware limitations, providing solutions to improve their workflow and achieve better image quality despite the slower processing speeds.

💡Ultimate Upscale Extension

The Ultimate Upscale Extension is a tool for the Automatic1111 software that allows users to upscale images by dividing them into tiles and applying settings to each tile. This method is highlighted in the video as a way to achieve higher quality images compared to the standard upscaling features within the software.

💡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 the image into 512 by 512 tiles, processes each tile individually, and then combines them to form the final upscaled image. This technique is shown to enhance the detail and quality of the upscaled image.

💡Denoising

Denoising is a process used in image and video processing to reduce noise and graininess, which can degrade the quality of the upscaled image. The video mentions adjusting the denoise strength as a critical setting in the upscaling process, where higher values can lead to more detailed images but might also alter the original content.

💡Realistic Vision Version 2

Realistic Vision Version 2 is a specific model used within the Automatic1111 software for image processing. The video uses this model as an example when discussing the upscaling process, emphasizing its role in achieving highly detailed and realistic results in the upscaled images.

💡ControlNet 1.1

ControlNet 1.1 is an updated version of a model used for image processing, which the video suggests using in conjunction with the Ultimate Upscale Extension. It is noted for providing more detail in the upscaled images, although it may also increase the rendering time due to its complexity.

💡Face Restore

Face Restore is a feature within the upscaling process that attempts to enhance the clarity and detail of faces in images. The video discusses turning this feature off when using tiled upscaling to avoid potential issues with facial details not rendering correctly.

💡PNG Info Tab

The PNG Info Tab is a part of the Automatic1111 interface where users can view and manage settings related to their images. The video script mentions using this tab to load images with their settings into the image-to-image tab for the upscaling process.

💡ESRGAN

ESRGAN, which stands for Enhanced Super-Resolution Generative Adversarial Networks, is a model used for upscaling images. The video discusses using the ESRGAN model within the Ultimate Upscale Extension to achieve high-quality results, especially for anime-style images.

💡Inpainting

Inpainting is a technique used to fill in or restore missing or damaged parts of an image. The video demonstrates using inpainting to manually enhance the eyes in an image after the upscaling process, which can result in a more natural and detailed appearance.

Highlights

The video introduces a method for high-quality upscaling on slow GPUs using an extension called Ultimate Upscale for Automatic 1111.

Ultimate Upscale splits images into 512x512 tiles, applying settings to each for higher quality results.

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

After installation, users should check for updates and restart the UI for the extension to take effect.

The video recommends using the 4X Ultra Sharp Upscaler for an additional upscaling method.

The 4X Ultra Sharp Upscaler can be downloaded from a Mega upload link and installed into the Automatic 1111 models folder.

When using the Ultimate Upscale, it's important to match the model used for the initial image creation with the upscaling model.

Adjustments are needed for tiling, such as removing specific image information and turning off the restore face feature.

Euler a with 20-30 steps is suggested for the sampling method, with 25 steps being suitable for most purposes.

Denoise strength between 0.2 and 0.3 is recommended for most cases, but higher values can yield more details.

For slow GPUs, the fastest method involves using the script with a scale from image size and an upscale to 4.

ControlNet 1.1 can be used for more detailed results, but it requires additional updates and model downloads.

Examples are provided to compare the quality of upscaling with different methods, including Realistic Vision version 2 and the 4X Ultra Sharp model.

Anime style images can be upscaled using the RESRGen 4X Anime model with the Ultimate Upscale script.

Face details can be improved post-upscaling by using a masking technique in a painting software.

ControlNet with tile can be used for higher denoise strength and more intricate details, though it requires more render time.

The video concludes with a side-by-side comparison of the different upscaling methods and their results.

The presenter encourages viewers to like the video and join the live stream for more content.