OpenOutpainter Serving Input Interrogate

openOutpaint has an "interrogate" tool - the A1111-era feature where you click an image and it tries to write a caption for it, usually so you can grab a prompt from something you pasted in. On the A1111 backend that went through its interrogate API; in this pack, the request lands at /sdapi/v1/interrogate and this node is what unpacks it.

The whole exchange is tiny: the request is {"image": "<base64>", "model": "clip"} - and the model field is hardcoded to "clip" by openOutpaint, so it's decorative. This node only cares about the image. It takes the one oop_request input and outputs a single IMAGE, decoded from base64 with OpenCV (that's the pack's opencv-python dependency earning its keep).

Wiring it

The important realization: this node does not caption anything. It just hands you the image. The actual captioning is up to your workflow - you plug the IMAGE into a captioning or tagging node of your choice (a CLIP interrogator, a WD14-style tagger, whatever you normally reach for), and then send the resulting text to the matching Output Interrogate node, which ships it back to the canvas. This is the pack's whole philosophy in miniature: the canvas asks, the workflow decides how to answer, the canvas gets the result.

Like the other input nodes, it's lazy and command-aware: it only fires when the pending request is an interrogate call, otherwise the IMAGE output comes back as an ExecutionBlocker and the branch idles. If your graph is captioning fine when you test it directly but doing nothing when the canvas calls it, that blocker is working as intended - you've wired it to a branch that isn't seeing an interrogate request.

Install and gotchas

Standard pack install, once for all nodes: ComfyUI Manager (search "OpenOutpaint ComfyUI Interface") or git clone https://github.com/without-ordinary/openoutpaint_comfyui_interface into ComfyUI/custom_nodes, then restart. opencv-python is the only pip dependency worth mentioning; no models to download for this node.

The gotcha is the captioning gap: if you wire this node straight to the Output node without a captioner in between, you'll be shipping an empty or wrong caption back. And the pack-wide rule applies - one serving workflow at a time, and if a graph error kills the branch before the output node finalizes, the canvas request hangs until you re-run with the same request_id.