Google Image Search

The name is a lie, and the good kind. This node doesn't call the Google API, doesn't need a key, and in fact has never seen Google. It's a DuckDuckGo image scraper wrapped in a ComfyUI node - type a phrase like "a golden retriever puppy," and it runs a live search, downloads a matching image into your input folder, and hands you a ready-to-use IMAGE tensor plus an alpha MASK. "Google Image Search" is marketing; the mechanism is duckduckgo_search. Don't let that put you off. For pulling reference photos into a workflow without manually saving files into ComfyUI/input, it does exactly what it says on the tin.

Why you'd reach for it

The README's example flows tell the real story: Google Image → Text (Caption) → Image. You search a reference image, run it through a captioning node, and feed the caption back as a prompt. The output is a genuine IMAGE tensor, so the same search result can also feed an img2img sampler, a ControlNet preprocessor, or an IPAdapter reference instead - reference-chasing without ever leaving the graph. It's a data node in ComfyUI's plumbing layer: no model, no pixels of its own, just a way to get outside content into the pipeline.

How it works

Under the hood it's two dependencies doing the work: duckduckgo_search runs the actual image query, and fastdownload pulls the file. The node requests up to three results (max_results=3), downloads each into ComfyUI/input/image_downloads/, then loads the first one that opens cleanly, skipping any that fail. The MASK output is the image's alpha channel, inverted - so a transparent PNG comes out with a usable mask - and zeros if there's no alpha.

One genuinely fun mechanism: every filter dropdown includes a Random option, and when you pick it the node shuffles that filter's choices and pops one until it lands on a real value. So you can lock your search phrase and let size, color, and region roll dice every run.

The inputs that matter

Every input has a default, so the node runs if you only touch search_phrase - that's the single field beginners actually set. The rest are DDG's native image filters exposed as dropdowns:

• search_phrase - what to search, multiline, defaults to "a golden retriever puppy."
• safesearch - on, moderate, or off (default off).
• size, time_period, color, type_image, layout, license_image, region - filters, each with Random. If you're building a licensed reference set, set license_image; otherwise leave them alone.

Outputs: image (IMAGE) wires into anything that takes an image, and mask (MASK) into masking/inpainting. Because it's an output node, it also emits a UI text box with the JSON of the first result - title, URL, and download path - useful for auditing what you actually got.

Installing it

ComfyUI Manager can find it, but the repo name has a typo - google-image-comyui-node (yes, "comyui") - so search "google-image" or by the author christian-byrne. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/christian-byrne/google-image-comyui-node
cd google-image-comyui-node
pip install -r requirements.txt

Then restart ComfyUI. The shipped requirements.txt installs duckduckgo_search>=7.0.0, fastcore, and fastdownload. Note the README says duckduckgo_search==5.2.1 but the actual requirements file asks for 7.0.0+ - trust the requirements file. If you hit a pip conflict, that version range is the first suspect.

Common issues

The one real gotcha is the author's own warning: 202 errors mean DuckDuckGo is rate-limiting you. This is a scraper, not an API, and DDG pushes back on bursts - the fix is a VPN or just waiting a while. It's inherent to the design, so don't expect a permanent cure.

Also worth knowing: downloads accumulate in input/image_downloads/ and never clean up after themselves, so delete that folder periodically. Each run is a live web call, so results change run to run, and on niche queries DDG can return nothing usable - the node just logs the failures and moves on. And with a three-result cap and one image out, this is a reference grabber, not a batch downloader.