ComfyUI Flux Detailer & Upscale

This is a workflow that I've had very good success with and wanted to share. I started with the wonderful workflow found here: https://civitai.com/models/698637/comfy-ui-flux-dev-with-upscaler-2-loras-and-adetailer-super-clean and adjusted it a bit. Props to this awesome workflow and if you are looking for a complete txt2img solution, give that one a look.

This workflow is not an end to end workflow for txt2img. It is an img2img workflow. That's one of the biggest differences. Use this workflow once you already have an image ready.

Second, while there are prompts involved here (and optional LoRAs) there are two separate processes. The first is the optional face detailer process. This comes with its very own prompt and optional LoRAs. The second is the upscaler process which is also a detailer as you are upscaling the latent and not just the pixels.

Keeping the prompts and LoRAs separate is intentional and important (though both still use the same base model of course). You may want to adjust a face in an image with a LoRA designed for face details, jewelry, tattoos (face tattoo anyone?), etc. while you use a completely separate prompt and LoRA(s) for the upscaler process.

This is an incredibly powerful way to upscale and enhance the details of your images.

I can't overstate just how powerful this is. If you've never experimented with upscaling the latent space, please give this workflow a try as it may be an eye opener for you.

I'll share any updates I make in the future, but I've been using this workflow for a while now and am pretty happy with it. I don't want to add more nodes to it. I want to keep it simple and focused on details and upscaling.

Just a Heads Up!

Because it doesn't just upscale the pixels, this takes a while to run. On my RTX 3090 Ti, it took between 10-12 minutes for just the upscale stage (not using the face detailer stage) for a 2x upscale on a 1920x1088 (2MP) Flux generated image. This makes it a 4k image and it looks great, but it takes some time. Your needs may not involve an upscale to as high of a resolution, but I noticed for Flux, you're better off starting from a 2MP image. It will upscale lower resolution images, but there just isn't a lot to work with there. You could run it through a second time though.

In fact, you may need to run this in multiple stages if you run out of memory. In other words, if it doesn't work at a 4x upscale, run two 2x upscale jobs.
I'm not entirely sure if you'll run out of memory or not, but it could take a really long time. I once accidentally set the input image to something that was already upscaled to 4k. It took me over an hour to realize my mistake, but it never crashed, I stopped the generation. So just be aware, this yields amazing results, but it can take some time.