ULTRA SHARP Upscale! - Don't miss this Method!!! / A1111 - NEW Model

Olivio Sarikas
23 Mar 202310:22

TLDRThis video introduces a method for achieving ultra sharp upscales using a specific model called 'four times Ultra sharp'. After downloading the model, the process involves rendering an image at high resolution with a denoise strength of 0.5, and then sending it to the 'Ultra sharp' app scaler for a further two times upscale. The result is a four times upscaled image with high resolution and sharp details. The video also explains the concept of latent image and why the method works, comparing it with a standard ESRGAN model. It highlights the superior coherence, texture, and detail of the 'Ultra sharp' model, especially in areas like hair, skin, and clothing. The video concludes with additional tips on using different sampling methods for image to image upscaling to maintain facial features and achieve sharper results.

Takeaways

  • 📚 Download the 'Four Times Ultra Sharp' model and place it in the 'models' folder within the 'ESRagain' folder.
  • 🖼️ When rendering an image, use the 'Hi-Res Fix' option and set the 'Upscale' to twice the original size with a 'Denoise Strength' of 0.5.
  • 🔄 After rendering a high-resolution image, send it to 'Extras' and select 'Ultra Sharp' from the 'App Scalers', setting it to 'Two Times' upscale.
  • 🔍 The process results in a four times upscaled image that is high resolution, super sharp, and rich in detail.
  • 🎨 The 'latent image' concept allows for better detail when upscaling, as the model can refer to latent data to enhance the image before it's fully rendered.
  • ⚙️ Turning off 'Hi-Res Fix' can save on GPU time and allow for faster rendering while still achieving a high-quality upscale.
  • 🔍 Experiment with 'Denoise Strength' values to find the right balance between image detail and noise reduction.
  • 📈 Using the 'Image to Image' upscale method allows for faster rendering and refinement before committing to a final upscale.
  • 🚀 The 'Ultra Sharp' model is preferred over the 'ESRagain' model for its consistency in texture and finer details, especially in areas like hair, skin, and clothing.
  • 🔧 Changing the sampling method, such as using 'DPN++ 2M Keras', can yield a more detailed and sharper result.
  • ⏱️ Upscaling in 'Image to Image' mode is more efficient for initial image refinement, saving time and resources before finalizing the upscale.

Q & A

  • What is the name of the model that is used for ultra sharp upscales?

    -The model used for ultra sharp upscales is called 'four times Ultra sharp'.

  • Where should the 'four times Ultra sharp' model be placed in the file system?

    -The 'four times Ultra sharp' model should be placed into the 'automatic 1111' folder, which is located in the 'models' folder within the 'ESR' folder.

  • How does the high-res fix feature work in the upscaling process?

    -The high-res fix feature first renders the image and then upscales it before it becomes an actual image, which helps to maintain more detail in the upscaled version.

  • What is the denoise strength setting recommended for both high-res fix and image to image upscale?

    -The recommended denoise strength setting for both high-res fix and image to image upscale is 0.5.

  • What is the difference between upscaling with high-res fix and image to image upscale?

    -High-res fix is more resource-intensive as it upscales the image before it's fully rendered, while image to image upscale is more economical as it allows you to render a low resolution image first, find a version you like, and then upscale it, saving on GPU time.

  • Why is the 'Ultra sharp' model preferred over the 'ESRGAN' model?

    -The 'Ultra sharp' model is preferred because it adds more texture to the image, provides finer details, and results in a more coherent and higher quality upscale.

  • How does the sampling method affect the final upscaled image?

    -The sampling method can significantly affect the sharpness and detail of the final image. Different methods like Euler or DPM++ 2m Keras can yield different results, with some being sharper and more detailed.

  • What is the benefit of using a lower denoise strength value like 0.25 in image to image upscaling?

    -Using a lower denoise strength value like 0.25 allows the AI to stick closer to the original image while still adding new details, resulting in a sharper image with more textures and details.

  • What is the purpose of the latent image in the upscaling process?

    -The latent image allows the upscaling algorithm to look at the latent data and add more detail to the upscaled image, resulting in higher quality and more details than a direct upscale of a low resolution image.

  • How does the quality of an image upscaled using the standard ESRGAN model compare to one upscaled using the Ultra sharp model?

    -An image upscaled using the standard ESRGAN model tends to look more like a digital drawing with less detail and coherence, whereas the Ultra sharp model provides a more consistent texture, finer details, and an overall higher quality upscale.

  • What is the recommended upscaling method when you are still searching for a good result?

    -When searching for a good result, it is recommended to use the image to image upscaling method as it allows for faster rendering and less resource usage before finding a satisfactory image.

  • Why might one choose DPM++ 2m Keras over Euler as the sampling method for upscaling?

    -One might choose DPM++ 2m Keras over Euler for a more detailed and sharper result, especially when working with photography or when the finer details and edges are important.

Outlines

00:00

📈 Upscaling Techniques for Ultra Sharp Images

The paragraph introduces a method for achieving ultra sharp upscales of images. It details a process involving the use of a specific model named 'four times Ultra sharp', which should be placed in the automatic 1111 folder within the ESR-GAN models directory. The method involves rendering an image at a high resolution with a denoise strength of 0.5, then sending it to an app scaler, specifically the 'Ultra sharp' scaler, set to two times the original size. This results in a four times upscaled image that is sharp and detailed. The paragraph also discusses the concept of a latent image and the importance of rendering before upscaling to preserve details. It concludes by comparing the quality of this method to a traditional upscaler, emphasizing the superior detail and sharpness of the former.

05:01

🔍 Deep Dive into Image Upscaling Quality

This paragraph delves deeper into the image upscaling process, contrasting the quality of upscaled images using the described method against the ESRGAN upscaler. It points out the deficiencies of the ESRGAN upscaler, such as the lack of detail in eyelashes, hair, and clothing, and a general low-quality appearance. The paragraph then explains why the 'Ultra sharp' model is preferable, demonstrating its ability to add texture, finer details, and a more coherent and realistic look to the upscaled images. It also discusses the importance of selecting the right sampling method for image to image upscaling to maintain facial features and details, mentioning different methods like Euler and DPM++ 2m Keras for different results. The paragraph concludes with a side-by-side comparison of the different upscaling methods, highlighting the improved results with the 'sde' method.

10:03

👋 Closing Remarks and Viewer Engagement

The final paragraph serves as a closing to the video script, thanking viewers for watching and encouraging them to like the video if they enjoyed it. It also prompts viewers to explore other content available to watch, suggesting that there is more interesting material to discover. The paragraph ends with a casual and friendly tone, expressing hope to see the viewers again and reminding them to leave a like if they haven't already.

Mindmap

Keywords

💡Ultra sharp upscale

The term 'Ultra sharp upscale' refers to a method of enhancing the resolution and clarity of an image to a very high degree. In the video, this process involves using a specific model to upscale the image four times, resulting in a high-resolution, sharp, and detailed output. It is central to the video's theme as it is the primary technique being discussed and demonstrated.

💡Model

In the context of the video, a 'model' is a software component used in AI applications for image processing. The 'four times Ultra sharp' model mentioned is a specific tool within the AI that helps in upscaling images. The model is crucial as it dictates the quality and style of the upscaled image.

💡Automatic 1111 folder

The 'Automatic 1111 folder' is likely a directory or folder within a software application where specific models are stored for use. In the video, the presenter instructs viewers to place the 'four times Ultra sharp' model in this folder for it to be utilized effectively. It is a key component in setting up the environment for the upscaling process.

💡Hi-res fix

'Hi-res fix' is a feature or setting within the image rendering process that allows for a higher resolution image to be produced. The video explains that selecting this option and using a two times upscale with a denoise strength of 0.5 is part of the method to achieve ultra sharp upscales.

💡Denoise strength

Denoise strength is a parameter that controls the level of noise reduction applied to an image during processing. A value of 0.5, as suggested in the video, strikes a balance between reducing noise and preserving image details. It plays a significant role in maintaining image quality during the upscale process.

💡Latent image

A 'latent image' is an image that is not yet fully formed or visible but contains the underlying data that can be used to create a detailed image. The video discusses how upscaling a latent image allows for more detail to be added to the final upscaled image, which is a key concept in achieving ultra sharp results.

💡Image to image upscale

This term refers to a process where an existing image is used as a base to create a higher resolution version. In the video, it is presented as an alternative method to the high-res fix, allowing for faster rendering and experimentation before committing to a final upscaled image.

💡ESRGAN (Enhanced Super-Resolution Generative Adversarial Network)

ESRGAN is a type of GAN (Generative Adversarial Network) used for super-resolution tasks, which is a technique for upscale images with improved quality. The video compares the results of using the 'Ultra sharp' model to those of the ESRGAN model, highlighting the differences in quality and detail.

💡Sampling method

The 'sampling method' is a technique used during the image upscaling process to determine how the new pixels are generated. Different methods can produce different results in terms of sharpness and detail. The video mentions Euler and DPM (Diffusion Probabilistic Models) as examples, with the latter being noted for producing sharper results.

💡Texture

In the context of image processing, 'texture' refers to the fine details and patterns that give an image a more realistic and three-dimensional appearance. The video emphasizes the importance of texture in the upscaled image, noting that the 'Ultra sharp' model adds texture and finer details to elements like hair and clothing.

💡Coherence

Coherence in image processing means the consistency and logical continuity of the details across an image. The video discusses how the 'Ultra sharp' model provides more coherence to the upscaled image, making it look more natural and less like a digital drawing.

Highlights

Download the 'four times Ultra sharp' model and place it in your automatic 1111 folder within the ESRGAN folder.

Render images at a high resolution by using the 'hi-res fix' and a two times upscale with a denoise strength of 0.5.

After rendering a high-res version, use the 'send to extras' feature and select 'ultra sharp' for an additional two times upscale to achieve a four times upscale result.

The final upscaled image will be high resolution, super sharp, and rich in detail.

Latent image upscaling allows for more detail by upscaling before the image is fully rendered.

An economic alternative is to render a low-resolution image, find a version you like, and then upscale using the image-to-image method.

Set the denoise strength to 0.5 when upscaling in the image-to-image method for better detail.

Experiment with the denoise strength value to achieve a balance between image detail and noise.

Upscaling with the Ultra sharp model after an initial two times upscale results in a high-quality, detailed image.

The quality of the upscaled image using the Ultra sharp model is identical to the high-res upscaled model, but rendered faster.

Compared to a standard four times upscale, the Ultra sharp model provides significantly better quality and detail.

The Ultra sharp model is superior to the ESRGAN model, offering more coherence, better fitting details, and sharper smaller elements.

The Ultra sharp model adds texture and finer details, especially at the ends of hair, enhancing the realism of the upscaled image.

Sampling methods such as Euler and DPM++ 2m Keras can be experimented with for different results, with DPM++ 2m Keras providing a slightly sharper result.

Using a lower denoise strength value like 0.25 with image-to-image upscaling can help maintain the original image's features while adding new details.

The Ultra sharp model is more consistent in bringing out finer details and providing a sharper, more detailed upscaled image.

When upscaling clothing details, the Ultra sharp model maintains texture and sharpness across the entire area, unlike the ESRGAN model which can be blurry.

The video provides a side-by-side comparison of different upscaling methods, showcasing the superior quality and detail of the Ultra sharp model.