How to AI Upscale and Restore images with Supir.

Sebastian Kamph
10 May 202416:31

TLDRThis tutorial demonstrates how to upscale and restore images using the Supir AI tool. The video guides viewers through installing necessary custom nodes in Comfy UI, selecting a suitable checkpoint and model, and adjusting parameters like the upscale factor and control scale for optimal results. Tips for fine-tuning images, such as altering prompts and sampler settings, are provided to achieve better restoration for low-quality or broken images. The workflow simplifies the process, making it accessible for users to enhance image details efficiently.


  • 😀 Install necessary custom nodes and restart Comfy UI if you encounter missing nodes and errors.
  • 🔍 If you don't have a local installation, use Think Diffusion, a cloud solution for running workflows without a powerful machine.
  • 🛠️ The workflow provided is a simplified version of the official one, focusing on ease of use rather than advanced features.
  • 📈 Use the 'upscale by X Factor' setting to determine how much larger you want the image to be compared to the original.
  • 🔍 For resource-intensive tasks, consider using the Lightning XL model, which is more efficient for image workflows.
  • 📚 Choose the right model for your needs from Civitai, which offers various options like Realistic and Epic Realism.
  • 🔄 Understand the difference between the Q and F models of Superar, with Q1 being the general-purpose model and F1 trained with light degradation settings.
  • 💬 Use prompts to guide the image restoration process, with positive prompts for desired details and negative prompts for undesired qualities.
  • 🎛️ Adjust the 'Super Control Scale' to find the right balance between retaining the original image details and allowing the AI to introduce new details.
  • 🔄 Experiment with different values for 'CFG Scale' and 'Super Control Scale' to fine-tune the image restoration process for various types of images.
  • 🖼️ For images with poor quality or broken details, fine-tuning the workflow settings and using manual prompting can significantly improve the restoration results.

Q & A

  • What is the main purpose of the video?

    -The main purpose of the video is to demonstrate how to upscale and restore images using an AI tool called Supir.

  • What is the first step in using the workflow for image upscaling and restoration?

    -The first step is to install the necessary custom nodes that are missing in the workflow by using the manager in Comfy UI.

  • Why might there be an error when trying to run the workflow?

    -An error might occur if custom nodes are missing, which can be resolved by installing them through the manager in Comfy UI.

  • What is the recommended way to find a Lightning XL model for upscaling?

    -The recommended way to find a Lightning XL model is by visiting Civitai, where there are various options such as Realistic and Epic Realism.

  • What are the two types of models mentioned in the video for using with the workflow?

    -The two types of models mentioned are the Q1, which is a general-purpose model, and the F1, which is trained with light degradation settings.

  • How can one adjust the level of detail and quality in the upscaled image?

    -One can adjust the level of detail and quality by modifying the 'super control scale' value and using appropriate prompts.

  • What is the significance of the 'CFG scale' and 'super control scale' in the workflow?

    -The 'CFG scale' and 'super control scale' are important for controlling the level of detail and coherence in the upscaled image, with lower values allowing for more freedom and creativity.

  • What does the video suggest for dealing with images that are not upscaling well?

    -The video suggests changing the sampler, adjusting the 'super control scale', and modifying the prompt to better describe the image content.

  • How can one ensure that the upscaled image retains the style of the original?

    -By using a higher 'super control scale' value, the upscaled image will retain more of the original style and details.

  • What is the recommended approach for images that are broken or have low quality?

    -For broken or low-quality images, the video recommends using a lower 'super control scale' value, adjusting the sampler, and providing a clear and specific prompt.

  • What are some alternative upscaling methods mentioned in the video?

    -Some alternative upscaling methods mentioned include using different resamplers and integrating other AI upscalers like ESRGAN.



🖼️ Image Restoration and Upscaling with Comfy UI

This paragraph introduces the process of restoring and upscaling images using Comfy UI. It explains the initial setup, including the installation of custom nodes and the use of a workflow file. The speaker addresses common errors and guides viewers on how to resolve them by installing missing custom nodes through the manager. The paragraph also touches on the use of checkpoints and the selection of appropriate models for the task, recommending resources like Civitai for finding suitable models. The workflow is described as user-friendly, with a focus on core functionality rather than advanced features.


🔧 Customizing Workflow Settings for Image Upscaling

The speaker discusses the customization of the workflow for image upscaling, including the use of prompts to guide the AI in generating images. They explain the importance of setting the correct values for control scale and CFG scale, which are crucial for the quality of the upscaled image. The paragraph also covers the use of different samplers and the impact of the seed value on image generation. The speaker provides examples of how altering these settings can affect the final image, emphasizing the need for fine-tuning to achieve the desired result.


🎨 Fine-Tuning Image Restoration with Superior Models

This section delves into the fine-tuning process for image restoration using superior models. The speaker provides a detailed walkthrough on adjusting the control scale to balance between preserving the original image details and allowing the AI to introduce new elements. They demonstrate the effects of different control scale values on a red truck image, showing how reducing the scale can improve the restoration quality. The paragraph also highlights the importance of using the correct prompt to guide the AI in generating a more accurate and detailed image.


📸 Testing Different Prompts and Settings for Upscaling

The final paragraph focuses on testing various prompts and settings for upscaling images. The speaker shares their experience with different prompts, such as 'elderly man cinematic photo', and how they affect the upscaled image. They also discuss the impact of changing the resampler and the potential for using AI-based upscalers like ESRGAN. The paragraph concludes with a demonstration of upscaling an image to a significantly larger size, showing the impressive results and the importance of using the right workflow for image restoration and upscaling.



💡AI Upscale

AI Upscale refers to the process of increasing the resolution of an image using artificial intelligence algorithms. In the context of the video, the AI upscale process is used to enhance the detail of images, making them appear clearer and more defined when enlarged. The script mentions using an AI model to upscale an image by a factor of two, which means the image's dimensions are doubled, resulting in a larger image with more detail.


In the video, 'restore' is used in the context of image enhancement, where AI is employed to improve the quality of old or degraded images. The script discusses using AI to restore details that may have been lost or degraded over time, such as in the example of the red truck, where the AI helps to bring back the clarity and detail of the image.


Superar appears to be a term used in the script to refer to a specific AI model or software used for image upscaling and restoration. The script mentions 'Superar' in the context of a workflow that simplifies the process of using AI for image enhancement, making it accessible and easy to use for users without extensive technical knowledge.

💡Comfy UI

Comfy UI is mentioned in the script as a user interface for running workflows related to AI image processing. It seems to be a platform where users can install custom nodes, manage models, and execute tasks like upscaling and restoring images. The script provides a tutorial on how to install and use Comfy UI for those who are new to it.

💡Custom Nodes

Custom Nodes in the script refer to additional components or plugins that can be installed in Comfy UI to extend its functionality. The video mentions that users may encounter errors due to missing custom nodes when they first run a workflow, and it guides them on how to install these missing nodes to proceed with the image processing tasks.


A checkpoint in the context of the video is a saved state of the AI model that can be loaded for specific tasks. The script talks about loading a checkpoint, such as 'lightning XL,' which is a model used in conjunction with Superar for image processing. Checkpoints allow users to utilize pre-trained models to perform complex image operations efficiently.

💡CFG Scale

CFG Scale, as mentioned in the script, refers to a configuration setting that affects how an AI model processes an image. It is used to control the degree of detail and coherence in the upscaled or restored image. The script explains that this value is usually lower for lightning models compared to regular stable diffusion models.


In the context of AI image processing, a 'prompt' is a text description that guides the AI in generating or modifying an image. The script discusses using positive and negative prompts to enhance the quality of the image, such as specifying 'high quality detail' to improve the outcome of the image processing.

💡Stable Diffusion

Stable Diffusion is an AI model mentioned in the script that is used for generating images from textual descriptions. While the main focus of the video is on upscaling and restoration, Stable Diffusion is brought up as a related technology that can be influenced by prompts to improve image quality.


A 'seed' in the context of AI image generation refers to a random number or set of numbers that initiates the process and influences the outcome. The script explains that the seed is randomized by default but can be set to a fixed value to reproduce the same image results consistently.


VRAM, or Video Random Access Memory, is the memory used by graphics processing units (GPUs) for storing image data. The script mentions settings related to VRAM usage, such as 'High VRAM' and 'FP8 unit,' which are options for users to adjust based on their hardware capabilities to manage resource-intensive tasks like AI image processing.


Introduction to AI image upscaling and restoration with Super.

Installation guide for Comfy UI and resolving common errors.

Managing missing custom nodes and installing them through the manager.

Using Think Diffusion as a cloud solution for those without a local machine.

Explanation of the simplified workflow designed for ease of use.

Importance of selecting the right checkpoint for the workflow.

Recommendations for finding and installing Lightning XL models.

Difference between Super models Q1 and F1 and their use cases.

How to adjust the prompt for better image generation results.

Adjusting control scale and CFG scale to fine-tune image detail.

Impact of sampler choice on resource usage and image quality.

Dealing with VRAM issues by toggling FP8 unit and High VRAM options.

Demonstration of image upscaling without changing any settings.

The effect of altering the Super control scale on image coherence.

Strategies for improving restoration of low-quality images.

Examples of how prompts can influence the outcome of image generation.

Fine-tuning image restoration with different control scale values.

The role of prompts in restoring images with specific features.

Experimenting with different prompts and settings for optimal results.

Upscaling high-resolution images and the impact on detail retention.

Customizing the workflow with different AI upscalers and resamplers.