IPAdapter Unified Loader V2

IP-Adapter needs three separate things before it can do anything: a CLIP vision encoder to read your reference image, the adapter weights themselves, and - for some presets - a companion LoRA. This node is the version of that setup where you don't have to think about any of it. Drop your base model in, pick a preset from a dropdown, and out comes the model plus an ipadapter pipe that every apply node in this pack consumes.

The node is a renamed copy of the Unified Loader from cubiq's ComfyUI_IPAdapter_plus (this whole pack is that project with V2 suffixes so it can run next to the original without node-name collisions). It figures out whether your checkpoint is SD 1.5 or SDXL by inspecting the model, then matches the file names against a list of known patterns. Pick PLUS (high strength) and it hunts for ip.adapter.plus.sd15 or plus.sdxl.vit.h depending on what you loaded. It even caches the weights so switching presets mid-session doesn't re-read the same file off disk.

The one input you actually set is preset. The choices tell you what they're for:

• LIGHT - SD1.5 only (low strength) - subtle nudge, biggest prompt freedom
• STANDARD (medium strength) - the original base adapter
• VIT-G (medium strength) - bigger CLIP encoder, better understanding
• PLUS (high strength) - patch-level tokens, tracks the reference much more closely
• PLUS FACE (portraits) and FULL FACE - SD1.5 only - face-region variants

Anything tagged "SD1.5 only" throws an error if you feed it an SDXL model, which is intentional - that's the node telling you the weights don't exist for your base, not a bug.

Both outputs matter. model is your checkpoint with any preset LoRA already merged in, and it feeds your sampler. ipadapter is a pipe carrying the loaded CLIP vision model, adapter, and (for FaceID) InsightFace handle, and it wires into IPAdapterV2, the Advanced, Tiled, or Precise nodes. You can thread an existing ipadapter pipe into the optional input to reuse a load instead of re-loading.

Install

This is a fork of cubiq's pack, GPL-3.0, with credit going back to Matteo "Matt3o/cubiq" Spinelli - support his original work. Two ways in:

• ComfyUI Manager: search ComfyUI_IPAdapter_plus_V2 and install
• or manually:

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

Restart ComfyUI. There's no requirements.txt - the pack leans on ComfyUI's own torch/torchvision/PIL, so no extra pip install for this node. Only FaceID presets need insightface (see the FaceID loader article).

Models you still have to download

The node won't fetch anything for you. Drop the weights where ComfyUI looks for them:

• ComfyUI/models/ipadapter/ - the adapter files (ip-adapter-plus_sd15.safetensors, etc.)
• ComfyUI/models/clip_vision/ - the CLIP vision encoder (CLIP-ViT-H-14-laion2B-s32B-b79K for SD 1.5, ViT-bigG-14 for SDXL)
• ComfyUI/models/loras/ - only needed for FaceID presets

Common gotchas

The classic failure is a preset you can't actually run: LIGHT on an SDXL checkpoint raises a clean "not supported for SDXL" error, and FULL FACE on SDXL does the same. If you get a terse "IPAdapter model not found" or "ClipVision model not found," the file name doesn't match any known pattern - check you have the right variant for your base and that it's in the exact folder above. And if a workflow loads with red "missing" nodes, it's usually because the original used cubiq's v1 names; this V2 pack deliberately renamed everything, so rewire to the V2 classes or install the original pack alongside.

One honest note: IP-Adapter is a legacy-SDXL-era tool. It's still the default way to do style and face conditioning on SD 1.5 and SDXL, but it doesn't load on Flux or newer architectures, and cubiq's upstream repo is in maintenance mode. If your workflow runs on SDXL and needs image prompting, this is the loader you'll reach for. If you're on something newer, this pack isn't your answer.