Dead-Simple Ultimate SD Upscale Workflow (Illustrious/Pony/SDXL)

I've tried to create the easiest Ultimate SD Upscale workflow possible.

SFW/NSFW, doesn't matter.

This workflow is intended for upscaling the already-awesome images you've created.

I looked around CivitAI for a workflow like this without spaghetti noodles. I guess I have a way of making workflows where all the nodes are close together in a box, kinda like an Automatic1111 or SwarmUI interface. Because once everything is configured correctly, why should you have to move around an enormous canvas? Anyway...

---

This workflow assumes you have a general understanding of what you're doing. That includes things like:

• Why you're using SD Ultimate Upscale in the first place

• What upscale models you like to use

• The checkpoint/LoRAs/sampling parameters you want to use

  • I usually use identical model/LoRA/cfg/sampler/scheduler as my original image with a 0.30 denoise. Positive prompt is just quality tags for me, like '8k, absurdres, masterpiece', but this depends on what you're aiming for.

• All of the parameters in the Ultimate SD Upscale node below the sampler/scheduler stuff( mask_blur, tile_padding, etc etc), are intended to be modified to get your desired output. What I have in the workflow is what works for my images but YOU need to tweak it and iterate.

---

What models does it work with?

This workflow is intended for any SDXL-based model that uses the standard checkpoint file you find on CivitAI. I use Illustrious exclusively but I don't see why it shouldn't completely work 100% with Pony, NoobAI, etc.

---

Custom nodes

As always, I try to use core ComfyUI nodes as much as possible to reduce bloat and annoyance. This workflow requires 4 custom node packs to work:

ComfyUI-Custom-Scripts (for the math expression nodes in the Backend subgraph):

https://github.com/pythongosssss/ComfyUI-Custom-Scripts

rgthree-comfy (for the awesome LoRA loader and image comparer):

https://github.com/rgthree/rgthree-comfy

ComfyUI_UltimateSDUpscale (for the upscaling node):

https://github.com/ssitu/ComfyUI_UltimateSDUpscale

ComfyUI_tinyterraNodes (for the positive and negative prompt nodes):

https://github.com/TinyTerra/ComfyUI_tinyterraNodes

---

How to use it

1. Drag any image onto “Load Input Image” (or choose a file).

2. Add LoRAs as needed.

3. Choose your upscale model (e.g., Real-ESRGAN, NMKD Siax, UltraSharp, etc.)

4. Change “Upscale By…” to 2 (default = 2×), 4 (4×), 1.5, 3, whatever you want. This node increments in steps of 0.5.

5. Add your positive and negative prompts.

6. In the Ultimate SD Upscale node, change around the steps/cfg/sampler/scheduler/denoise as you wish.

7. Hit Queue Prompt → Done.

What you get instantly:

• An upscaled image that respects your original aspect ratio exactly.

• Saves the final upscaled image straight to your output folder.

• Built-in side-by-side slider (rgthree Image Comparer) so you can see before/after instantly.

---

What happens behind the scenes

There’s one tiny orange “Backend” box that does all the magic automatically:

• Detects your input resolution

• Multiplies width & height by your upscale factor → calculates exact target dimensions

• Feeds those target dimensions into proper SDXL-style conditioning (positive + negative)

• Gives the tiling sampler perfect knowledge of the final size it’s aiming for

Feel free to open or expand the subgraph as needed.