📷InsightFace Loader

If you're building an InstantID workflow with this pack, this is the InsightFace loader you actually want - not its lookalike twin. It shows up in the node list as "📷InsightFace Loader," same as the other one, which is a genuinely confusing bit of naming. The tell is the output type: this one produces INSIGHTFACEMODEL, and that's the exact type the InstantID Generation node's insightface input expects.

How it works

Same job as any InsightFace loader - detect a face, align it, turn it into an ArcFace identity embedding - but this is Zho's own build of the loader rather than the one borrowed from cubiq's ComfyUI_IPAdapter_plus. The pack's v2.0 changelog says plainly it exists to fix a naming conflict: install this pack alongside anything else that also registers a node literally called InsightFaceLoader (cubiq's own IPAdapter pack, for instance, since that's where the code was lifted from) and ComfyUI has to resolve which one wins. Renaming the class to InsightFaceLoader_Zho - and giving it its own output type - sidesteps that collision entirely and makes sure this pack's InstantID Generation node always gets exactly the object it's built for.

The input and output that matter

One setting: provider, CUDA or CPU. CUDA is the obvious pick if you've got the hardware - this model is cheap to run relative to everything else in an InstantID graph, so it's rarely the bottleneck either way. Output is INSIGHTFACEMODEL, which wires directly into the insightface input on InstantID Generation - nowhere else in this pack consumes it.

Installing it

Search ComfyUI Manager for ComfyUI-InstantID, or clone it by hand: cd ComfyUI/custom_nodes && git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-InstantID.git, then cd in, pip install -r requirements.txt, restart. That install pulls in insightface, which is a genuinely rough dependency to get working - it needed a source build on older Python versions before InsightFace shipped its 1.0 release. If your install is stuck compiling, that's why.

The model this loader needs - antelopev2 - isn't auto-downloaded. Get it from DIAMONIK7777/antelopev2 and place the whole thing under ComfyUI/custom_nodes/ComfyUI-InstantID/models/antelopev2. Don't substitute the buffalo_l pack InsightFace pulls automatically - InstantID's IdentityNet was trained against antelopev2's specific detector/recognizer combo, and it's the pack most InstantID tools quietly assume you already have.

Common issues

"Insightface" errors that trace back to a missing or misplaced model, not this node itself. This is the most common failure mode across the whole InstantID-in-ComfyUI space - the node loads fine, then errors deep inside a face-analysis call because antelopev2 isn't where it's expected. Re-verify the exact path before chasing anything else.

Grabbed the wrong InsightFace Loader. If you're getting a type mismatch wiring into InstantID Generation's insightface slot, you very likely picked the plain InsightFaceLoader instead - check the node's class name (visible in its properties panel), not just its display title, since both read "📷InsightFace Loader" at a glance.

CUDA 12 users specifically. The bundled onnxruntime-gpu defaults to a CUDA 11 build (1.16.0). On CUDA 12 you need to manually install onnxruntime-gpu==1.17.0, or this loader (and anything else touching InsightFace in the pack) can fail or silently downgrade to CPU.

Commercial use. The antelopev2 weights this node loads are non-commercial research models. That restriction is inherited by every InstantID output downstream of it, independent of the wrapper code's own license.