Load Upscale Model (Online)

This is the node that makes the whole pack make sense. Instead of scrounging a Dropbox link for 4x-UltraSharp and hand-placing it in models/upscale_models, Load Upscale Model (Online) lists the upscalers DoubTech hosts, downloads the one you pick on first use, and hands it to ComfyUI's stock upscaler loader. It's a convenience wrapper, not a new way to upscale.

The pack's upscaler catalog is small but it's the good stuff: 4x Ultra Sharp, 4x NMKD-Siax (the anime/illustration pick), 4x Foolhardy, and x1 ITF Skin Diff Detail Lite. These are the classic ESRGAN-family pixel upscalers - the "more pixels, adds no content" rung of the upscaling ladder, per the KB's upscaling essay. If you're expecting a generative restorer like SUPIR or SeedVR2 that invents detail, this isn't it. It's the fast, predictable, can't-hallucinate pass you run before a hires-fix second sampling step or as the final output upscale.

How it works

Under the hood it's the same deal as every other DT node in this pack: on ComfyUI startup the pack fetches the upscaler list from DoubTech's API (api.aiart.doubtech.com/comfyui/upscalers), and when you queue the node it downloads your selection into your local models/upscale_models folder - unless a file with that name is already there, in which case it just loads it. No API key, no hosted inference. "Online" means "on-demand download from a shared catalog." Once a file lands locally it's yours; the node never touches the network for it again.

The one input and what it wires into

The only input is model_name, a dropdown populated from the catalog. The single output is UPSCALE_MODEL. Plug it into the core ImageUpscaleWithModel node, then either feed the upscaled pixels back into a low-denoise KSampler (the classic hi-res fix) or straight into your save/encode step.

Install

Everything in this pack installs the same way. In ComfyUI Manager, search "DTAIComfyLoaders" (the pack title is "Comfy UI Online Loaders") and hit install, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/yolanother/DTAIComfyLoaders

Then restart ComfyUI. The only dependency is requests, which you almost certainly already have - there are no heavy installs hiding in this pack. It's MIT-licensed, written by Yolan (DoubTech), and last touched in late 2024.

Where people get burned

• First use downloads the model. 4x-UltraSharp is a ~65 MB file, so it's quick by model standards, but there's no progress bar in the UI. The node just sits there on the first run.
• A "downloaded" model that fails to load. Several files in DoubTech's catalog live on Dropbox, and Dropbox serves scripted downloads an HTML bot-check page rather than the raw file. If your upscaler errors out on load, check the file size in models/upscale_models - if it's a few hundred KB of HTML instead of tens of MB, the download was the problem, not the node. Grab the file from the link manually and drop it in the folder.
• The dropdown is empty on a fresh restart. The catalog fetch happens in a background thread; give it a moment and re-add the node. And there's no refresh button - the list only updates when ComfyUI restarts, so new catalog entries take a restart to appear.

Honestly? If you already keep your upscalers tidy locally, the stock UpscaleModelLoader does the same job. This node is for the person who just wants 4x-UltraSharp without the treasure hunt.