SDXS: 90 millisecond renders with comfyUI

Stephan Tual
3 Apr 202407:21

TLDRThe video discusses the challenges of slow loading times in KY's interface due to numerous custom nodes and models. The speaker presents a solution by reinstalling KY and using a clean install to significantly improve rendering speeds, achieving 90 millisecond renders with SDXS technology. The process involves disabling previews in the manager, using the verbose command line parameter for transparency, and configuring the model path in the YAML file. The result is a streamlined workflow that allows for faster image generation, even with a less powerful system.

Takeaways

  • 🚀 The video discusses improving the render times and performance of a complex KY (Kerchief) setup with many custom nodes and models.
  • 🔍 The speaker identifies that the slow loading times are due to the high number of custom nodes and additional notes running in the background.
  • 🛠️ A potential quick fix is to disable previews in the manager, but this is more of a temporary hack and not a permanent solution.
  • 🎯 The speaker decides to kill the current server instance and download a fresh copy of KY, emphasizing the importance of a clean install for improved performance.
  • 📋 The video provides a step-by-step guide on how to install a new version of KY, including unzipping and placing it into a designated folder.
  • 🌟 An important tip is given to add '--verbose' at the end of the Nvidia GPU bat file to help with debugging and understanding what KY is doing.
  • 📂 Instructions are given to modify the 'extra model path.yml' file to point the new KY install to the folder containing the existing models.
  • 🔄 After starting the server with the new setup, it is noted that the load times are significantly improved, with the interface loading in microseconds.
  • 🖼️ The video demonstrates the use of SDXS (an imaging technology) to generate images quickly, with the new setup allowing for faster rendering of images.
  • 📈 Comparing the old and new installs, the video shows that the new install with fewer folders and a clean setup can generate images in 0.09 to 0.10 seconds, a notable improvement.
  • 🚀 The video concludes by encouraging viewers to experiment with the new workflow, even on less powerful computers, for improved performance and faster render times.

Q & A

  • What is the main issue discussed in the transcript?

    -The main issue discussed is the slow loading time and performance problems of an existing Confy UI installation, specifically in the browser, which takes up to 5 seconds to load the GUI.

  • What is the suggested solution to the performance problem?

    -The suggested solution is to kill the current server instance, download a fresh copy of Ky, and properly configure the new installation to use the existing model files from the old installation.

  • How can one identify the function causing performance issues?

    -One can use developer tools to select the function that's currently causing trouble, which can help identify the source of the performance issues.

  • What is the role of SDXS in the transcript?

    -SDXS is used as an example of a technology that can quickly generate images, albeit with a lower quality. It demonstrates the potential speed improvements after resolving the performance issues.

  • What is the significance of disabling previews in the manager?

    -Disabling previews can be a temporary workaround to improve performance, but it is not a permanent solution as the previews are still useful for users.

  • How does the speaker propose to find the path for the models used by Ky?

    -The speaker suggests using the command line parameter '--verbose' to have Ky output the paths it's using for models, which can help in troubleshooting path-related issues.

  • What changes were made to the 'extra model path.yml.example' file?

    -The '.example' part of the file extension was removed to convert it into a regular yaml file. The speaker then uncommented lines of code and specified the correct base path where the existing models are located.

  • What was the result of the new Ky installation after the changes?

    -The new Ky installation loaded instantaneously compared to the regular version, and it was able to find and use the specified model paths, resulting in improved performance and rendering times.

  • How much faster is the new Ky installation compared to the old one?

    -The new Ky installation reduced the image generation time from an average of 0.19 to 0.20 seconds in the old version to 0.09 to 0.10 seconds in the new version, achieving nearly 10 frames per second.

  • What is the benefit of using an empty latent image batch?

    -Using an empty latent image batch allows for processing 100 images at once, which can significantly increase the speed of image generation, although the quality may still be compromised.

  • What is the final recommendation for users with performance issues?

    -The final recommendation is to follow the outlined workflow, which involves a fresh installation of Ky and proper configuration to utilize existing model files, ensuring improved performance and usability.

Outlines

00:00

🚀 Optimizing Interface Performance and Workflow Efficiency

The paragraph discusses the challenges of dealing with a slow-loading interface due to an accumulation of models and custom nodes over time. The speaker shares their personal experience with an existing confy install that takes a long time to load, particularly in the browser, and the impact it has on both interface and rendering performance. They provide a solution by suggesting the use of developer tools to identify problematic functions and offer a step-by-step guide on how to refresh the Ky server instance and configure it to use the existing model paths, resulting in significantly improved performance and efficiency. The speaker emphasizes the importance of a clean install and proper configuration to avoid performance issues and maintain a smooth workflow.

05:01

🎨 Harnessing the Power of SD XS for Rapid Image Generation

This paragraph focuses on the utilization of SD XS, a technology that uses one-step diffusion to quickly generate images, albeit with variable quality. The speaker demonstrates how to adjust the settings to improve the speed of image generation, such as changing the batch size and enabling auto queue. They also explain how to dynamically change prompts to generate different types of images, showcasing the technology's flexibility and speed. The speaker compares the performance of their new, optimized setup with their old one, highlighting the significant reduction in image generation time. The paragraph concludes with a recommendation to batch process images for even greater speed and a mention of the workflow being available for others to experiment with, regardless of their computing power.

Mindmap

Keywords

💡SDXS

SDXS refers to a technology mentioned in the video that uses a one-step diffusion process to rapidly generate images. Although the quality of the images produced may not be the highest, the speed at which they are generated is impressive, allowing for a high volume of outputs in a short amount of time. In the context of the video, the user is utilizing SDXS to produce a large number of images quickly, highlighting the technology's efficiency in rendering images, albeit with a trade-off in quality.

💡KY

KY, as mentioned in the video, appears to be a software or platform that the user is working with. The user experiences slow loading times with KY, particularly in the browser, which can take up to 5 seconds just to load the GUI. This indicates that KY is a significant part of the user's workflow, but its performance issues are causing problems. The user seeks a solution to improve KY's performance, which is a central issue addressed in the video.

💡confyUI

confyUI is referenced as an existing installation that the user is dissatisfied with due to its slow loading times. The user describes it as taking a long time to load and negatively impacting the performance of both the interface and the rendering process. To resolve this, the user decides to download a fresh copy of confyUI and reinstall it, which is a key step in the solution provided to improve the overall performance of the system.

💡performance

Performance in this context refers to how well and how quickly the software or technology under discussion (KY, confyUI, SDXS) can execute tasks or processes. The user is concerned with improving performance, particularly the loading and rendering times, which are affected by various factors such as the number of custom nodes and the amount of data being processed. The video aims to provide solutions to enhance performance, making it a central theme of the discussion.

💡custom nodes

Custom nodes are user-created components or extensions that add specific functionalities to the software they are used with. In the video, the user mentions having a high number of custom nodes running, which is contributing to the performance issues experienced with KY and confyUI. The user's situation is representative of scenarios where customization can lead to complications, and the video explores ways to manage these issues.

💡previews

Previews in this context refer to the visual representations or snapshots of the models or images within the software. The user discusses disabling previews as a temporary workaround to improve performance, suggesting that these previews, while useful, can slow down the system. This highlights the trade-off between utility and performance that users may face when working with complex software environments.

💡server instance

A server instance is an individual, virtual or physical server that runs a specific application or service. In the video, the user decides to kill the current server instance and start a new one as part of the solution to improve performance. This action signifies a complete restart of the service, which can help eliminate issues stemming from accumulated data or processes in the previous instance.

💡Nvidia GPU

Nvidia GPU refers to a Graphics Processing Unit manufactured by Nvidia, a company known for its high-performance graphics and computing products. GPUs are essential for rendering images and videos, as they can handle complex calculations required for graphics processing more efficiently than general-purpose CPUs. In the video, the user provides a tip on configuring the Nvidia GPU settings to enhance the software's performance, indicating the importance of leveraging hardware capabilities to optimize software operations.

💡yaml file

A yaml file is a human-readable data serialization format often used for configuration files. In the video, the user edits a yaml file to specify the model paths for the software, which is a crucial step in ensuring that the newly installed software can access the existing models. This demonstrates the importance of properly configuring the software environment to maintain the workflow and utilize existing resources effectively.

💡autoq

Autoq, as used in the video, appears to be a feature or command that automates the queuing process for tasks, such as generating images. The user enables autoq to speed up the workflow, allowing for a more efficient and faster operation. This highlights the use of automation tools in streamlining processes and enhancing productivity within complex software environments.

💡prompt

In the context of the video, a prompt is a text input or a command given to the software to generate specific types of images. The user changes the prompt to 'a cute dog' and then to 'a cute cat on a mountain with wings,' demonstrating how the software responds to these inputs by generating images that match the descriptions. This showcases the dynamic nature of the software and its ability to create images based on textual instructions.

Highlights

The discussion focuses on optimizing the rendering process with comfyUI, addressing common issues faced by users with extensive model collections.

A significant problem highlighted is the slow loading times of the GUI for KY, which can take up to 5 seconds, affecting both interface and rendering performance.

The use of developer tools to identify and disable problematic functions, such as CJs linked to custom nodes, is mentioned as a temporary solution.

SD XS is introduced as a technology that rapidly generates images using a one-step diffusion process, albeit with variable quality.

Disabling previews in the manager is suggested as a quick fix for performance issues, but it is not ideal as previews are useful.

The speaker proposes a permanent solution by reinstalling KY and leveraging the benefits of a clean install to improve performance.

A detailed guide on reinstalling KY is provided, including downloading the latest version and setting up the configuration correctly.

The importance of specifying the correct model path in the config UI folder is emphasized to ensure the new install recognizes existing models.

The use of the 'verbose' command line parameter is recommended to gain insights into KY's operations and troubleshoot potential issues.

The new install's performance is demonstrated to be significantly faster, with a reduction in image generation time from 0.19-0.20 seconds to 0.09-0.10 seconds.

Batch processing is suggested for even faster rendering, with the potential to reach 100 frames per second under optimal conditions.

The analogy of cookies in ovens is used to explain the efficiency of batch processing versus single task processing.

The workflow shared in the transcript is available on float, allowing users to experiment with the methods discussed, regardless of their hardware capabilities.

The speaker emphasizes the value of the discussed methods in optimizing performance, especially for those with extensive custom node usage.

The transcript concludes with a call to action for viewers to try out the workflow and explore the potential of the discussed techniques.