External Image (ComfyDeploy)

External Image (ComfyDeploy) is the interesting member of the ihmily/comfy-deploy "External" family, because it's the only one that does real work in the node itself. The text, int, and float versions are pass-throughs - this one fetches bytes. It's the node that lets an API caller hand your deployed workflow an image: input for img2img, a ControlNet reference, an IPAdapter image, an inpainting mask source.

The two-part system is the same as the rest of the pack - nodes in ComfyUI, plus a self-hosted admin backend (comfy-deploy-admin) that manages and publishes deployed workflows. In the editor you feed this node an image via its default_value and test normally. When a caller's request includes your param_name, the admin injects that value - a URL or a base64 data URL - and the node downloads and decodes it at runtime.

How it works

Unlike the other External nodes, load_image has real branches. Look at nodes/comfydeploy_external_image.py and the logic is:

• Value starts with http → fetched with httpx using a browser user-agent, 3 retries, a 60-second timeout, following redirects. This is the normal case for an API call.
• Value is a base64 data URL (data:image/png;base64,...) → decoded locally. No network round-trip, which is what you want if the caller's image lives inside their own app.
• Anything else → falls back to default_value.

The fetched image gets EXIF-transposed (so phone photos aren't sideways), converted to RGB, normalized to a 0–1 float tensor, and returned as an IMAGE batch of one. If the fetch fails outright, the node raises a RuntimeError - it will not silently proceed.

Inputs and outputs that matter

• param_name (required, default "input_image") - the request key. Same story as the rest of the family: this is your API contract, keep it stable.
• keep_alpha_channel (required, default off) - the one setting that affects output, not just naming. Leave it off for normal RGB image work; turn it on when you genuinely need transparency, like feeding an RGBA image into an inpainting or background-removal stage.
• default_value (optional, an actual IMAGE) - your local-testing image and the fallback when a call omits the parameter. You can pick it like any image input in the editor.

The output is image (IMAGE) - wire it into LoadImage-style inputs: img2img, ControlNet's reference port, IPAdapter, anything that takes an image.

Installing it

Ships with the pack, so one install gets you all five External nodes:

cd ComfyUI/custom_nodes/
git clone https://github.com/ihmily/comfy-deploy.git
cd comfy-deploy
pip install -r requirements.txt

Or via ComfyUI Manager (search "comfy-deploy", author ihmily). No model files. README gate: pack v1.0.1+ needs ComfyUI 0.3.67+. Then set up the admin backend and point the ComfyDeploy panel at it (endpoint + API key) before deployment works.

Where people get burned

The URL you supply must be reachable from the ComfyUI server, not from your browser. http://localhost:3000/myimage.png on your dev machine is unreachable from a server across the internet - this is the classic "works when I test, fails in production" image-API bug. Either host the image somewhere public (an object store, a signed URL) or pass a base64 data URL instead. And note the URL only gets fetched on the deployed path or when you type it into the widget - the default_value image you set in the editor is what flows through locally.