IPAdapter Unified Loader (Clone)

Anyone who's set up IPAdapter by hand knows the ritual: which CLIP Vision model goes with which adapter file, for which SD version, did you rename the download, is the filename going to match what the loader regexes for... IPAdapter Unified Loader (Clone) exists to end that ritual. It's the pack's copy of cubiq's IPAdapter Unified Loader from ComfyUI_IPAdapter_plus - the "(Clone)" in the name isn't a branding quirk, it's a licensing acknowledgment. Pseudocomfy's README credits Matteo Spinelli's pack directly, and the code is a trimmed port.

Pick a preset, feed it your model, and it returns your model unchanged plus an ipadapter bundle (CLIP Vision + adapter, ready for any IPAdapter conditioning node). The preset dropdown in the published node lists six options: LIGHT - SD1.5 only (low strength), STANDARD (medium strength), VIT-G (medium strength), PLUS (high strength), PLUS FACE (portraits), and FULL FACE - SD1.5 only (portraits stronger). If you're on SDXL and want the pack's recommended setup, PLUS is your preset and it maps to the ip-adapter-plus_sdxl_vit-h.safetensors model the README tells you to download.

How it decides what to load

This is the part worth understanding, because it's where the "just works" can break. Under the hood it does filename pattern matching against what's actually in your models folders:

• It detects your base model. It checks whether the model you wired in is an SDXL family model and picks adapter/CLIP-Vision patterns accordingly - so one preset dropdown serves both SD 1.5 and SDXL.
• It regexes models/clip_vision and models/ipadapter for filenames matching the preset (ip.adapter.sd15, ip.adapter.sdxl.vit.h, plus.sdxl.vit.h, and so on). If nothing matches, it raises a plain "IPAdapter model not found." That's your #1 failure mode, and it's almost always a naming or folder problem, not a code problem.
• It can load a companion LoRA for plus-style presets when one is present in models/loras, applying it to the model - that pairing is part of how the "plus" models get their extra strength.
• InsightFace loading is deliberately disabled - the author raised an exception stub in place of the FaceID path. So FACE presets degrade to CLIP-based face conditioning rather than the research-only InsightFace route. That's a real licensing-aware choice: the CLIP-based variants stay Apache-2.0 and commercially usable.

Practical setup

You don't install anything extra for the loader itself - but you absolutely must have the weights. The README's checklist: cubiq's ComfyUI_IPAdapter_plus nodes (or just the model files), a CLIP Vision model in models/clip_vision, and an IP-Adapter .safetensors in a models/ipadapter folder you may need to create. The recommended downloads for SD 1.5 / SDXL are ip-adapter-plus_sd15.safetensors and ip-adapter-plus_sdxl_vit-h.safetensors. Keep the standard filenames - the loader matches on them.

There's also an optional ipadapter input so you can chain or override an existing bundle, which matters when you run several of these loaders with different presets in one graph. The output ipadapter wires straight into PseudoApplyIPAdaperSDXL (or any IPAdapter Apply node). If the community IPAdapter consensus holds true, keep weights moderate on SDXL - Plus at 0.6–0.8, not 1.0 - and you'll get reference steering without the reference eating your prompt.