Load Upscale Model (URL/Path)

This one does exactly one thing: you give it a URL (or a local path), and it hands you a loaded UPSCALE_MODEL you can plug straight into an "Upscale Image (using Model)" node. That's the whole job - it's the URL-fetching twin of ComfyUI's built-in Load Upscale Model node, which only ever looks inside your local models/upscale_models folder.

It's part of ComfyUI-load-lora-from-url, a personal utility-node pack by bollerdominik that, despite the name, has grown well past just LoRAs - a video loader, a couple of mask-compositing helpers, and a handful of API-backed image/video generation nodes all live in the same repo. The pack has no real footprint anywhere online and the README is a single sentence, so there's nothing pack-specific to lean on beyond what the node itself does.

Why you'd want this instead of the stock loader

Upscale models - the ESRGAN family, mostly: 4x-UltraSharp, Remacri, the NMKD variants - are non-generative pixel upscalers. They add no new content, can't hallucinate, and run in milliseconds. Reliable, boring, and exactly what you want when the source image already has the detail and you just need it bigger. The catch is you have to actually own the .pth file first: download it from OpenModelDB or wherever, drop it in the right folder, restart ComfyUI so it shows up in the dropdown.

That manual step is annoying on your own desktop and outright impossible on a lot of cloud or shared setups where you can't just browse to a folder and drop a file in. This node sidesteps it - one url field, paste the direct link to the upscale model file, and it resolves the model at run time instead of asking you to pre-stage it.

The one input, and the one output

• url (required, multiline string) - a direct link to the upscale model file, or a local path if you already have it downloaded somewhere. That's the entire input surface.
• upscale_model - the loaded model, output type UPSCALE_MODEL. Wire it into the standard "Upscale Image (using Model)" node exactly as if you'd loaded it the normal way; nothing downstream needs to know it came from a URL.

Installing it

ComfyUI Manager: search ComfyUI-load-lora-from-url. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/bollerdominik/ComfyUI-load-lora-from-url

Restart ComfyUI. No extra dependencies are documented for this node specifically, and there's nothing to pre-download - that's the point of it.

Where people get burned

• Pointing it at a model page instead of a file. A link to a CivitAI or OpenModelDB page isn't the same as the raw download URL for the .pth itself. Grab the direct file link, not the page you'd browse to find it.
• Picking the wrong model for the job. These pixel upscalers are for "the image is already sharp, I just need it bigger." If your source is soft, blurry, or genuinely damaged, an ESRGAN-family model won't invent detail that isn't there - that's a different job (SeedVR2, SUPIR, tiled diffusion) and a different kind of node entirely. Using the wrong category is the most common reason people come away unimpressed by upscaling in general.
• Re-downloading every run if nothing caches it. The node fetches from the URL at run time; if your setup doesn't persist the download between runs, expect the fetch to happen again each time you queue the workflow. Worth checking once rather than assuming.