I was tired if having to constantly switch nodes around when upscaling, so I made this all-in-one, easy to use workflow to cover all of my upscaling needs.
I figured someone might find this useful as well, so I polished it up a bit, and here it is :)

TL;DR:

This is an all-in-one, multi-pass upscaling and detailing workflow for SDXL (incl. Illustrious, NoobAI, etc) and Flux.

It puts plenty of powerful nodes and techniques at your fingertips, allowing you to upscale an image as much as you want, adding details while retaining the original composition and avoiding deformation/doubling.

There are many settings that allow you to choose exactly how much detail to add, and how similar to the original image the output will be.

It's "plug and play", and you won't need to bypass or reroute any nodes. Everything is controlled via toggles and sliders.

I consider this workflow fairly beginner friendly, as you don't need to mess around with any nodes, but the more knowledge you have, the better you'll be able to wield this workflow.

PS: It also works great for img2img, with or without upscaling!

MAIN FEATURES:

Multi-pass Upscaling:

You can choose in how many passes the original image will be upscaled to the target factor, scaling linearly and ensuring that each pass increases the resolution by the same amount of pixels. Many parameters have a "First Pass" and "Last Pass" value, that will scale the values throughout the passes. This allows for a high degree of flexibility, letting you choose when to give more freedom to the model, and when to restrain it to prevent oddities or unwanted changes.

Composition Retention:

You have many tools at your disposal to help retain the composition of the original image, such as: A ControlNet (with up to 3 preprocessors at the same time), ClownGuide, the Unsample/Resample technique, scaling values, and even more. These allow you to use very high Denoise values while upscaling, which is useful to add new, coherent details while still maintaining the original composition.

Detail Injection:

You can add as much or as little detail to the picture as you like, both big and small. There is a plethora of settings to inject noise into the image that, when used in conjunction with the other parameters, will allow you to fine-tune the details of your output. At higher resolutions, you can add extremely fine details.

Note: The model and LoRAs you use can have a significant impact on the type and level of detail that can be achieved.

Tiled Upscale:

Tiled upscaling can be toggled on and off, and is fully integrated into this workflow using the Tiled Diffusion nodes.
All of the settings also work with tiled upscaling, but make sure to use a tiled ControlNet!

Latent & Model Upscale:

By default, all upscaling is done to the latent, only encoding and decoding the image once, preserving quality. There is also the option to use an upscale model, although you lose the benefits of latent upscaling, as upscale models cannot work on latents.

Ease of Use:

Everything is controlled via UI nodes, with sliders and toggles. You won't need to hop around the workflow to tweak settings, as everything is grouped in a control panel, without the need to bypass or reroute anything. Just load your image, tweak the settings as desired, and you're good to go.

Also, every option works with every other option! You can safely use any combination of settings without breaking the workflow (unless I missed a bug somewhere).

Dual Phase & Switch Passes:

The workflow uses two phases, both of which can be toggled on or off independently, allowing you to completely switch settings after a certain number of passes. For example, you can upscale normally in Phase 1, and then use Tiled Upscaling in Phase 2.

There is also the option to enable a "Switch Pass" during each phase, that allows you to change sampler, scheduler, and/or multiply many parameters during certain passes.

IMG2IMG:

This workflow can also be used for img2img, with or without upscaling. By allowing the model enough freedom, and changing your prompt accordingly, you can alter the style of the image, or change any number of things within it. You can also use a different model than the original image's to completely change style.

Check out the example images at the top of the page to see a couple of examples of what can be done.

Monitoring & Previews:

There is a "Live Monitor" that updates at every pass, showing the values in use for the current pass. On top of that, there are image previews for the following things: Latest Pass, Phase 1 Result, Phase 2 Result, Final Result (after post-processing), ControlNet (blending every preprocessor in use into a single preview image).
Each preview will update throughout the workflow's execution when opportune.

GGUF Support:

Both .safetensors and .GGUF models are fully supported, including the VAE and CLIP.

Offloading to Secondary Device:

You can offload the CLIP, VAE, or even the main model to the CPU (or a second GPU), freeing VRAM.

RES4LYF:

This workflow uses the ClownShark samplers, and employs some powerful RES4LYF nodes, allowing you to tweak the sampling process in many meaningful ways.

Documentation:

Below every section you'll find a note, briefly explaining the settings and how they may affect your output.
If you're unsure about the function of something, you'll probably find the answer in those notes.