AI Models From ANY Angle & SUPIR Upscaling (Low VRAM Method)

Aiconomist
2 Jun 202414:23

TLDRIn this tutorial, viewers learn to create and upscale AI-generated images of a digital model from various angles while maintaining consistency. The video demonstrates using a Level Four XEL model with Comfy UI and Superpier for upscaling, ensuring high-quality results with minimal VRAM. Techniques include adjusting prompts for facial features, using depth control nets, and comparing different upscaler models for optimal image quality. The tutorial also covers separating images, refining character consistency, and leveraging Superpier for powerful image restoration and upscaling.

Takeaways

  • 🖼️ The video demonstrates how to create AI digital model images from various angles with consistency and quality.
  • 🚀 The 'Level Four XL' turbo model is used for its speed and versatility in generating realistic, anime, and CGI images.
  • 📊 Recommended settings for Level Four XL include 8-16 steps and a CFG scale of 1-3, with the DMP Plus+ SDS sampler.
  • 🎨 Comfy UI is used to load a workflow for generating character faces with consistent facial features from multiple angles.
  • 🔍 The 'Depth Anything Tool' and 'Apply Control Net Advanced' nodes are utilized for generating faces in specified directions.
  • 🖼️ Four different upscaler models are tested for image quality improvement using 'Ultimate SD Upscale'.
  • 📏 The 'Image Crop Node' is used to separate the upscaled image into four individual images, each showing the character from different angles.
  • 😄 The 'IP Adapter' and 'Unified Loader' are employed to generate a batch of images with consistent faces, clothing, and backgrounds.
  • 🔄 Weight adjustments in the 'IP Adapter' and 'Plus Face' can help achieve a balance between reference likeness and emotional expression.
  • 🛠️ 'Face ID' can be used alongside 'Plus Face' for improved facial feature accuracy, but it's not allowed for commercial use without permission.
  • 🌟 'Superpier' is highlighted as a powerful tool for image restoration and upscaling, capable of producing high-quality 2K and 8K images.

Q & A

  • What is the purpose of today's video?

    -The purpose of today's video is to teach viewers how to generate high-quality images of an AI digital model from different angles while maintaining consistency, and to upscale images using the SuperP image restorer and upscaler.

  • What type of GPU was used to create the content in the video?

    -An RTX 360 with 12 GB of VRAM was used to create the content in the video, from generating images to upscaling them.

  • What is the name of the checkpoint model used in the video?

    -The checkpoint model used in the video is called Level Four XL, which is a turbo model capable of generating images quickly and creating various types of visuals.

  • What are the recommended settings for the Level Four XL model?

    -The recommended settings for the Level Four XL model are between 8 and 16 steps with a CFG scale between 1 and 3, using the DMP Plus+ SDS sampler and setting the Schuler to caras.

  • How can viewers find the Level Four XL model?

    -Viewers can find the Level Four XL model on Civid AI.

  • What is the challenge presented in the video regarding the character's face?

    -The challenge is to generate a character's face from multiple angles while keeping the facial features consistent.

  • What tool is used to generate character faces in specified directions?

    -The Depth Anything tool is used in conjunction with an Apply Control Net Advanced node and the Zoe depth model for sdxl to generate character faces in specified directions.

  • What is the purpose of using the Ultimate SD Upscale node?

    -The Ultimate SD Upscale node is used to upscale images for better quality, and the video tests four different upscaler models to compare and choose the best results.

  • How can viewers access the upscaling models mentioned in the video?

    -Viewers can access the upscaling models by downloading them for free from OpenM Modelb Doino, where they can also see live examples and additional information about each model.

  • What is the recommended method for separating the upscaled image into four individual images?

    -The recommended method is to use the Image Crop node to crop the upscaled image into four individual images, each showing the character from different angles.

  • What is the role of the IP adapter in generating consistent character faces?

    -The IP adapter, when connected with models like Face ID and Plus Face, helps in generating consistent character faces with specific features, emotions, and angles as desired.

  • What is the SuperPier tool used for in the video?

    -SuperPier is used for restoring and upscaling images using text prompts, which is particularly useful for enhancing image quality and texture while maintaining the likeness of the characters.

  • What are the system requirements for using SuperPier smoothly?

    -For using SuperPier smoothly, especially with the right setup and models, you would need at least 12 GB of VRAM or more.

  • How long did the entire upscaling process take using a 12 GB GPU card?

    -Using a 12 GB GPU card, the entire upscaling process took 3 minutes and 20 seconds to generate a 2K and 8K image.

Outlines

00:00

🖼️ Generating Consistent AI Digital Model Images

This paragraph introduces the video's focus on creating high-quality images of an AI digital model from various angles with consistency. The process involves using an RTX 360 GPU for generating images with the IP adapter and upscaling them with Super, an image restorer and upscaler. The tutorial covers the use of the Level Four XL checkpoint model, a versatile hybrid for realistic, anime, and 3D art. It details the steps to set up Comfy UI, select the model, and use prompts to vary facial features while maintaining consistency. The recommended settings for the model and the process of generating character faces using a depth control net are also explained.

05:01

📸 Upscaling and Refining AI Model Images

The second paragraph delves into the upscaling process using the ultimate SD upscale node with different models to enhance image quality. It emphasizes the importance of refining and experimenting with the process to achieve the best results. The tutorial guides through separating the upscaled image into individual images using the image crop node and adjusting their positions. It also discusses the use of the IP adapter with the plus face portrait model to generate consistent character images with varying emotions and backgrounds. The paragraph highlights the impact of different weight values on the likeness to reference images and suggests using additional processors like Face ID for improved results.

10:08

🚀 Advanced Image Upscaling with SuperPier

The final paragraph introduces SuperPier, a powerful tool for image restoration and upscaling that requires significant GPU power. It provides a step-by-step guide on setting up the workflow, including installing necessary nodes and models, and using SuperPier for upscaling images to 2K and 8K resolutions. The tutorial explains the use of Juggernaut XL version 9 for optimal results and the importance of selecting the right model and settings for the upscale process. It concludes with a demonstration of the improved image quality and texture achieved through SuperPier, showcasing the facial features and likeness restoration without losing detail.

Mindmap

Keywords

💡AI Digital model

An AI Digital model refers to a computer-generated representation of a person or character, which can be manipulated to create images from various angles and with different features. In the video, the AI Digital model is used to generate consistent character faces from multiple perspectives, showcasing the versatility of AI in creating realistic or stylized imagery.

💡Upscaling

Upscaling is the process of increasing the resolution of an image or video, typically to enhance its quality or to prepare it for larger displays. The video discusses using the 'Superp' tool to upscale images of the AI Digital model, demonstrating how higher resolution can improve the detail and clarity of the generated images.

💡Checkpoint model

A checkpoint model in the context of AI refers to a specific state or version of a neural network that has been saved and can be reloaded for continued training or for generating outputs. The 'Level Four XEL' mentioned in the script is a checkpoint model used for generating images quickly and with various styles.

💡CFG scale

CFG scale, or Classifier-Free Guidance scale, is a parameter used in some AI image generation models to control the balance between the detail of the image and the adherence to the style or content specified by the user. In the video, it's mentioned that the recommended CFG scale for the Level Four XEL model is between 1 and 3.

💡DMP Plus+ SDS sampler

The DMP Plus+ SDS sampler is a specific type of sampler used in AI image generation that helps in creating high-quality images by sampling from the distribution of the generated data more effectively. The script instructs to use this sampler with the Level Four XEL model for optimal results.

💡Image crop node

An image crop node is a tool used in image editing software to select and extract a portion of an image. In the video, it's used to separate a single upscaled image into four individual images, each showing a character face from a different angle.

💡Ultimate SD upscale

Ultimate SD upscale refers to a method or tool used to increase the resolution of an image to Super Definition (SD) levels. The video script mentions testing different upscaler models to compare and choose the one that provides the best results for enhancing image quality.

💡IP adapter

An IP adapter in AI image generation is a tool that helps in injecting specific features or characteristics into the generated images based on the input provided. The script describes using the IP adapter with various models to create consistent character faces with specific emotions and backgrounds.

💡Face ID

Face ID, in the context of the video, refers to a feature or tool used to identify and manipulate facial features in AI-generated images. It's mentioned as a method to improve the consistency and likeness of the character's face across different images, although it's noted that it's not allowed for commercial use.

💡Superpier

Superpier is a term used in the video to describe a powerful image upscaling tool that can restore and enhance images using text prompts. It requires significant GPU power but can produce high-quality results, as demonstrated by the improved texture and facial features in the upscaled images.

Highlights

Learn how to create AI Digital model images from various angles with consistency.

Upscale images using the state-of-the-art image restorer and upscaler, SuperP.

The video was created using an RTX 360 with 12 GB of VRAM for generating and upscaling images.

Explore the checkpoint model 'Level Four XEL', a turbo model for fast image generation.

Discover how to generate character faces with consistent facial features from multiple angles.

Use descriptions like 'Spanish' or 'German' to vary facial features in the AI model.

Specify age to create a young or older character in the AI model.

Utilize the DMP Plus+ SDS sampler and set the Schuler to 'caras' for optimal results.

Create a single image with four character faces from different angles using a reference image.

Employ the 'Depth Anything' tool and connect it to the 'Apply Control Net Advanced' node.

Experiment with prompts to adjust features or fix issues in generated images.

Compare four different upscaler models to determine the best final result.

Download upscaling models from OpenM Modelb Doino for free and use them in Comfy UI.

Separate the upscaled image into four individual images using the 'Image Crop' node.

Generate a batch of four images with consistent faces, clothing, and backgrounds using the IP adapter.

Adjust the weight value to improve the similarity of the AI model's face to the reference images.

Use 'Face ID' along with 'Plus Face' for better facial feature consistency, but note its commercial use restrictions.

Explore the 'SuperPier' method for upscaling images with text prompts and its GPU power requirements.

Install necessary nodes and models for the SuperPier workflow to upscale images to 2K and 8K resolutions.

Experience the improved image quality and texture with SuperPier's upscaling capabilities.