IPAdapter Model Loader V2

IPAdapter Model Loader V2 is the explicit way to load an IP-Adapter weight file. One input - ipadapter_file, a dropdown of every adapter in ComfyUI/models/ipadapter - and one IPADAPTER output that you wire into the ipadapter socket of any apply node in this pack. It's the "I know exactly which weights I want" loader, and it's the piece you reach for when you'd rather control the pipeline node by node than let a unified loader decide for you.

How it fits

The alternative is IPAdapter Unified Loader V2, which picks the adapter, the CLIP vision model, and (for FaceID) the insightface model all at once based on a preset. That's convenient and most tutorials use it. But the unified loader's preset decides for you - and the more you know about your weights, the more you want the explicit path: IPAdapter Model Loader V2 for the adapter, IPAdapter Unified Loader... no, actually, a CLIP vision loader (or the clip_vision input on the apply node) for the vision model, IPAdapter InsightFace Loader V2 for face models when needed, and then an apply node. You pick the exact file instead of a preset's regex pattern finding it for you.

That file-selection detail is the practical difference. The unified loader matches filenames against a pattern per preset, and if your weights are named oddly it silently fails to find them. The Model Loader just shows you the whole folder. If you're juggling multiple adapter files for different jobs - Plus for style, FaceID Plus v2 for faces, a community Kolors adapter - this is the loader that lets you swap them without rethinking the graph.

What to put in the folder

The dropdown lists ComfyUI/models/ipadapter. Adapter files are small (the Plus adapters are around 100MB - 22M parameters - versus multi-GB checkpoints), and the ones this pack expects follow the Tencent/h94 naming: ip-adapter-plus_sd15.safetensors, ip-adapter-plus_sdxl_vit-h.safetensors, the FaceID line, and so on. You'll also need the matching CLIP vision model in ComfyUI/models/clip_vision because the apply nodes require it.

Installing it

ComfyUI Manager: search "ComfyUI_IPAdapter_plus_V2". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart ComfyUI, then drop the adapter files in models/ipadapter.

Gotchas

The classic error: an apply node wired with an IPADAPTER that has no CLIP vision attached and no clip_vision input filled → "Missing CLIPVision model." The Model Loader only loads the adapter; the vision model is on you. And note this pack is a renamed snapshot of cubiq's V2 codebase, so its file-matching behavior is pinned to that state - if upstream fixes a matching bug, this fork won't have it until it's synced.