How to UPSCALE with Stable Diffusion. The BEST approaches.

The paragraph introduces the limitations of StableDiffusion XL, such as its high memory requirement and common training resolutions of 512x512 or 768x768. It contrasts this with the improved performance and lower memory usage of custom models for StableDiffusion 1.5, which offer comparable or better image quality. The main challenge with these models is their lower resolution output, which the video aims to address through upscaling techniques. The paragraph also mentions the availability of these models on platforms like CivitAI and HuggingFace, and provides a brief guide on installing the epicRealism model and an upscaler called Superscale. The process of generating an image using the epicRealism model is outlined, including the need to adjust parameters such as sampling steps, height, and CFG scale.

This paragraph delves into the details of upscaling an image generated by a custom model. It advises caution when using certain prompts due to the potential for the model to generate images with insufficient clothing details. The paragraph discusses the use of the high-res fix as an upscaler, with options to choose different upscale modules like ESRGAN. It also touches on the importance of setting the denoise strength and the limitations faced when working with limited GPU memory. The narrative then shifts to a simpler upscaling method, resizing the image by 4 times using the nearest pixel method and ESRGAN for better results. It concludes with a discussion on using the image-to-image tab for further upscaling, emphasizing the need to adjust parameters and prompts for optimal results.

The paragraph focuses on advanced upscaling techniques using the SD upscale script and the Superscale upscaler. It emphasizes the importance of setting the correct parameters, including the sampling method, sampling steps, CFG, and denoising levels. The Superscale upscaler is chosen for its ability to generate more detailed and less pixelated images. The paragraph also discusses the benefits of using tile sizes that are compatible with the default settings of Stable Diffusion 1.5, which allows for efficient use of VRAM and the generation of upscaled images even on graphics cards with lower memory. The selection of the Superscale upscaler and the generation of the upscaled picture concludes the section, highlighting the improved detail and quality of the upscaled image.

This section introduces ControlNet as a superior upscaling solution, which is a neural network capable of various tasks, including defining a person's pose. The paragraph outlines the installation process of ControlNet, which is similar to that of the ultimate upscaler. It also discusses the necessity of downloading specific models from HuggingFace and placing them in the appropriate directory for upscaling. The narrative then guides through the process of using ControlNet for upscaling, emphasizing the stability of the results and the potential for repeated upscaling without significant quality loss. The paragraph concludes with a comparison of the results obtained from using ControlNet and the ultimate SD upscale script, highlighting the increased detail and stability of the ControlNet output.

The final paragraph summarizes the different upscaling methods discussed in the video. It recommends using ESRGAN for quick results without the need for detail, the SD upscale script or the ultimate SD upscale script for more detailed images, and ControlNet for the best results or when multiple upscaling iterations are required. The paragraph also provides a visual comparison of the upscaling results with and without the use of a script, emphasizing the superior detail retention and stability of the ControlNet method. It concludes with a call to action, encouraging viewers to like or comment if the video was helpful for upscaling their images.