Load Image from URL

Paste a URL, get an image tensor. That's the whole pitch of Load Image from URL, and it's genuinely handy if you feed reference images from the web into img2img, ControlNet, or IP-Adapter workflows. Normally you'd download the file, drag it into the ComfyUI input area, and only then reference it by filename. This node skips all of that - the URL is the input.

It lives under DoubTech/Image rather than the pack's usual DoubTech/Loaders category, because it's not an online model loader at all; it's a remote image fetcher. The default URL is even a QR code image from doubttech.ai, which tells you something about the author's taste - the pack also ships a qrcode ControlNet in its catalog.

Inputs and outputs

The only input is image, a plain string field holding the URL (multiline off). The default is https://www.doubtech.ai/img/doubtech.ai-qrcode.png. Outputs are the usual pair:

• IMAGE - the RGB tensor, normalized 0–1, ready to plug into a KSampler's image input or any image processor.
• MASK - built from the image's alpha channel (inverted, so opaque becomes 0/masked-out... actually inverted so the transparent area is the mask). If the image has no alpha channel, you get a blank 64×64 mask instead.

How it works

The node fetches the URL, runs an EXIF-transpose so phone photos don't come in rotated, converts to RGB, and normalizes to a tensor. Two details worth knowing:

• It supports data: URLs - base64-embedded images. That means you can pipe in an image that never touched disk at all, which is a neat trick for automation.
• It refetches every run. The node's change-detection always reports "changed," so it hits the network on every queue instead of caching. For a one-off that's fine; inside a loop that repeats a remote fetch, it's wasteful.

Install

Same as the whole pack - in ComfyUI Manager search "DTAIComfyLoaders" (title "Comfy UI Online Loaders"), or:

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

Restart after. The pack's only dependency is requests.

Where people get burned

• Hotlinking is fragile. The node does a plain GET with no browser headers, so URL-shortened links, signed URLs, and anything behind a bot check will fail or return an HTML page instead of pixels. If you get an error or a "image" that's not an image, fetch the file in a browser first and either host it somewhere direct or use the data: URL trick.
• No progress or timeout UI. A slow link just looks like a hung node.
• The default QR image. Don't forget to replace it - a surprising number of people queue the default and wonder why their img2img source is a QR code.

The core LoadImage node is better for files already in your input folder; this one shines exactly when the image lives on the web and you don't want to manage a file for it.