8K Upscaling Mastery with Stable Diffusion & Auto 1111: A No-Fluff Workflow Walkthrough

AIchemy with Xerophayze
9 Feb 202419:10

TLDRIn this video, Eric from alchem O, zero phase revisits the topic of upscaling images using a Stable Diffusion model. He emphasizes the importance of saving images as PNG to avoid compression artifacts and ensure better color depth for upscaling. Eric demonstrates his workflow using a portrait as an example, detailing the process of selecting the right model and settings for a high-quality upscale. He explains the use of prompts for generating detailed images, such as a street portrait, and the adjustments made during the upscaling process, including the use of DPM Plus+ 2m SD carass for more intricate details. The video also covers the use of high-resolution fix, denoise strength, and the Ultimate SD upscale method for a clean and detailed upscale. Eric cautions against using the extras tab for upscaling due to the loss of detail and recommends keeping the denoise strength low for subsequent upscales. The result is a detailed and realistic upscaled image, with a focus on texture and fine details, showcasing the effectiveness of the described method.

Takeaways

  • 🖼️ Save upscaled images as PNG to avoid compression artifacts and preserve color depth.
  • 🔍 Use DPM Plus+ 2m SD caras for more intricate details in realistic images.
  • 📐 Start with a narrower width for portraits and a 4x5 aspect ratio for better detail in skin, lips, eyes, and hair.
  • ⏱️ Keep sampling steps consistent and use hres fix to improve image quality without changing the seed.
  • 🔍 Use a detailer for faces to enhance details, but it's not always necessary if the initial render is good.
  • 📝 Generate prompts for specific styles, such as street portrait photography, to guide the AI's output.
  • 🧩 Use the Ultimate SD upscale method for high-quality upscaling without losing intricate details.
  • 🚫 Avoid using the extras tab for upscaling as it can smooth out details and lead to a loss of image integrity.
  • 🔄 For second-generation upscales (e.g., 4K to 8K), reduce the denoise strength to prevent unwanted patterns and maintain image integrity.
  • 🎨 After upscaling, additional realism can be added through inpainting with an inpainting model for detailed areas like faces.
  • ⏲️ The process is designed to be efficient, with the tutorial aiming to be under 20 minutes for quick learning.

Q & A

  • What is the recommended file format for upscaling images to avoid artifacting and banding?

    -The recommended file format for upscaling is PNG because it has a deeper color depth and no compression artifacts, which can cause issues like artifacting and banding when using JPEG.

  • Why does Eric prefer using the DPM Plus+ 2m SD model over the UL model for upscaling?

    -Eric prefers the DPM Plus+ 2m SD model because it tends to provide more intricate details, which is suitable for creating realistic images, unlike the UL model which has a smoother looking generation render.

  • What aspect ratio does Eric set for the portrait upscaling?

    -Eric sets the aspect ratio to 4x5 for the portrait upscaling.

  • Why does Eric suggest not using negative prompts when working with a well-trained SDXL model?

    -Eric suggests not using negative prompts with a well-trained SDXL model unless structural integrity is required, like with buildings, because the model is capable of generating high-quality images without them.

  • How does Eric approach generating prompts for upscaling street portrait photography?

    -Eric generates random prompts for street portrait photography by specifying the theme as 'Street portrait' and selecting prompts that include a confident urban woman and a charismatic street musician.

  • What is the purpose of using 'hres fix' during the upscaling process?

    -The 'hres fix' is used to add detail to the image during the upscaling process, ensuring that the high-resolution fix retains the quality and sharpness of the image.

  • Why does Eric recommend enabling a detailer when working with faces?

    -Enabling a detailer when working with faces helps to add more detail and clarity to facial features, such as the eyes, pupils, and hair, which can be important for achieving a realistic look in the final image.

  • What is the significance of using 'control net' and 'Pixel Perfect' during the upscale process?

    -Using 'control net' and 'Pixel Perfect' helps to ensure that the upscaled image maintains its integrity and detail, preventing issues like checkering patterns and preserving the overall quality of the image.

  • Why does Eric advise against using the 'extras tab' for upscaling?

    -Eric advises against using the 'extras tab' for upscaling because it can result in a loss of intricate details and a 'bleeding' effect, where the image appears smoothed out and loses its sharpness.

  • What is the recommended Doo strength setting when performing a second generation upscale from 4K to 8K?

    -The recommended Doo strength setting for a second generation upscale from 4K to 8K is between 0.15 to 0.2 to avoid creating unwanted checkered patterns and to maintain image integrity.

  • How does Eric suggest adding more realism to the upscaled image, especially regarding facial details?

    -Eric suggests using an inpainting model to add more realism to the upscaled image, particularly for facial details, by manually detailing elements like the skin texture, hair, and other features to enhance the realism.

Outlines

00:00

🎨 Upscaling Workflow and PNG Settings

Eric from Alchem O begins by discussing the importance of upscaling settings, particularly the use of PNG over JPEG to avoid compression artifacts and maintain color depth for smoother gradients. He emphasizes setting the saving format to PNG for upscaled images. Eric then demonstrates creating a portrait with intricate details in skin, lips, eyes, and hair using an SDXL model. He explains the choice of DPM Plus+ 2M SD Carass for its ability to render more detailed and realistic images, and outlines the process of generating prompts for street portrait photography.

05:00

🔍 Refining Upscaling Parameters

The second paragraph delves into the technical aspects of upscaling, such as keeping sampling steps consistent, enabling hres fix, and adjusting Doo strength for detail enhancement. Eric also mentions disabling the filer text box and using the same model for the refiner to avoid long wait times. He highlights the importance of controlling the level of change in each segment during upscaling and enabling control net with Pixel Perfect for a tiled upscale. The paragraph concludes with a focus on the Ultimate SD upscale method for achieving high-quality results without losing intricate details.

10:01

🚫 Avoiding the Extras Tab for Upscaling

Eric cautions against using the extras tab for upscaling, as it can lead to a loss of detail and a 'bleeding' effect that smooths out the image. He advocates for using the Forex ultr sharp upscaler instead, which he considers one of the best available. The paragraph also covers the process of setting the Doo strength, enabling control net, and using tile blur for a tiled upscale. Eric demonstrates the step-by-step upscale process and the importance of adjusting settings to maintain image integrity.

15:03

🔄 Double-Checking Upscaling Results and Avoiding Checker Patterns

In the final paragraph, Eric discusses the results of the upscaling process, noting the improved texture and detail in various elements of the image. He warns against increasing the Doo scale too much during a second-generation upscale, as it can introduce unwanted checker patterns and image integrity issues. Eric demonstrates the impact of different Doo strength settings and provides guidance on achieving the best results. He concludes by suggesting further realism can be added through inpainting with an inpainting model and invites viewers to subscribe for more tutorials.

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 the primary focus, where the host demonstrates how to enhance the quality of an image from a lower to a higher resolution without significant loss of detail or introduction of artifacts. It is crucial for creating high-definition images from lower quality originals.

💡PNG

PNG stands for Portable Network Graphics, which is a file format used for storing images on the web. The host emphasizes the importance of saving images as PNG rather than JPEG during upscaling. This is because PNG offers a deeper color depth and lacks compression artifacts, which helps maintain the quality of the upscaled image, especially important for preserving fine details and gradients.

💡JPEG

JPEG is a commonly used method of lossy compression for digital images. The video mentions that upscaling JPEG images can result in artifacting and banding due to its lossy compression, which reduces the quality of the upscaled image. This is why the host recommends using PNG for upscaling to avoid these issues.

💡SDXL Model

SDXL Model refers to a specific type of model used in the upscaling process. The host mentions using an SDXL model, which is likely a machine learning model designed to handle image upscaling tasks. The SDXL model is chosen for its ability to generate high-quality images with detailed textures and features.

💡DPM Plus+ 2M SD Carass

DPM Plus+ 2M SD Carass is a term used in the video to describe a particular setting or model variant used in the upscaling process. It is suggested to provide more intricate details and is suitable for creating realistic images. The host switches to this model to enhance the detail level in the upscaled images.

💡Aspect Ratio

Aspect ratio is the proportional relationship between the width and the height of an image or screen. In the video, the host sets the aspect ratio to 4x5, which is a common aspect ratio for portrait-oriented images. This setting is important for maintaining the correct proportions when upscaling.

💡Prompts

Prompts are the textual descriptions or cues that guide the generation of images in AI models. The host generates prompts for 'street portrait photography' to instruct the AI model on the style and content of the image to be upscaled. These prompts are critical for achieving the desired outcome in the final upscaled image.

💡Highres Fix

Highres Fix is a process mentioned in the video that seems to be part of the upscaling workflow. It is used to enhance the resolution of the image before further upscaling. The host suggests turning on Highres Fix to ensure the image quality is improved before proceeding with additional upscaling steps.

💡Control Net

Control Net is a feature or tool within the upscaling process that is used to manage and refine the details of the upscaled image. The host enables Control Net with 'Pixel Perfect' to ensure that the upscaled image maintains its integrity and detail, particularly when dealing with complex textures and patterns.

💡Ultimate SD Upscale

Ultimate SD Upscale is a method or tool used for upscaling images. The host discusses using Ultimate SD Upscale as one of the best ways to upscale an image without losing quality, as opposed to other methods that may smooth out details. It is part of the final steps in the upscaling process to achieve high-quality results.

💡Denoising Strength

Denoising Strength is a parameter that controls the level of noise reduction applied during the upscaling process. The host adjusts the denoising strength to add detail to the upscaled image without introducing unwanted artifacts. It is a critical setting when fine-tuning the upscaled image to achieve the desired level of detail and clarity.

Highlights

When upscaling images, it's important to save as PNG instead of JPEG to avoid artifacting and banding.

Using a normal SDXL model for upscaling can provide smoother looking renders.

Switching to DPM Plus+ 2M SD Cross Sampler can enhance intricate details for more realistic results.

Setting a narrower width for portraits ensures better detail in facial features.

For well-trained SDXL models, negative prompts are rarely needed unless structural integrity is required.

Generating random prompts for photography can provide a range of options to work with.

Street portrait photography prompts can help generate images with a subject in focus.

Upscaling can bring out details like texturing on clothing, skin texture, and hair details.

The hands in AI-generated images may appear mangled but can be overlooked if not the focus.

Image-to-Image upscale using PGS files can be slower due to their size but retains more detail.

ControlNet with Pixel Perfect and tile blur can be used for a tiled upscale without losing image integrity.

Doo strength should be carefully adjusted to avoid checker patterns in the upscaled image.

Ultimate SD upscale is recommended for high-quality image upscaling without losing intricate details.

Overlapping sections during the upscale process ensure a seamless blend of details.

Upscaling can be performed multiple times for higher resolutions, but Doo strength should be reduced for subsequent upscales.

Inpainting with an inpainting model can add more realism and detail to the upscaled image.

The presenter suggests that this upscaling method is the best until more advanced techniques are learned.

The video provides a quick and straightforward workflow for upscaling images without unnecessary complexity.