IPAdapter Model Helper (Inspire)

The genuinely miserable part of setting up IP-Adapter isn't the concept - it's the file matching. An IP-Adapter needs the right adapter weights and the right CLIP Vision encoder and, for the FaceID line, an InsightFace model and a companion LoRA, and every one of those has to match your base model and match each other. Get the CLIP Vision wrong (ViT-H vs ViT-bigG) and you get garbage or a cryptic shape-mismatch error. This node's entire reason to exist is to collapse that mess into a single dropdown.

You pick a preset - SD1.5 Plus, SDXL Plus Face ViT-H, Kolors Plus, one of 24 - and it loads the correct constellation of files for you, wires in the companion LoRA where the preset needs one, and spits out everything downstream in one bundle. If you've ever lost an evening to "why does IP-Adapter throw a tensor error," this is the node that makes it stop.

How it works

It's a smart loader. Each preset is a known-good recipe: adapter model + CLIP Vision + (for FaceID) InsightFace + the required LoRA at the right strengths, all matched to a specific base. It applies the LoRA to your model and clip, loads the encoders, and packages the adapter-side pieces into an IPADAPTER_PIPE so you don't have to hand-wire five separate loaders. Under the hood it's driving cubiq's ComfyUI IPAdapter Plus - Inspire is the convenience layer, not a reimplementation.

Reality check from the knowledge base: IP-Adapter itself is frozen. Tencent shipped its last variant in January 2024 and cubiq's node pack is in maintenance-only mode. It still owns image-prompting inside SD1.5 and SDXL workflows, but nothing here loads on Flux or newer - the presets top out at the architectures IP-Adapter ever supported.

The inputs and outputs that matter

The two you actually think about:

• preset - the whole point. Pick the one matching your base and the job (plain style, Plus for tighter tracking, Plus Face / Full Face for commercial-safe faces, FaceID variants for identity). One heads-up from the licensing side: anything with FaceID reads InsightFace embeddings and is research-only, so the plain and Plus Face presets are the ones that stay commercially usable.
• model (and optional clip) - your base model, which the node patches with the FaceID LoRA when the preset calls for one.

Also worth knowing: insightface_provider (CPU / CUDA / ROCM - set it to match your GPU for FaceID presets), insightface_model_name (buffalo_l or antelopev2), and cache_mode, which keeps the InsightFace and/or CLIP Vision models resident so you're not reloading them every run.

Outputs are deliberately generous: IPADAPTER_PIPE (the tidy bundle - feed it to Regional IPAdapter nodes or FromIPAdapterPipe), plus the unpacked IPADAPTER, CLIP_VISION, INSIGHTFACE, and the patched MODEL and CLIP, and two cache-key strings. Use the pipe for cleanliness, or grab the individual wires if a node wants them directly.

How to install it

ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

then restart. It's a Dr.Lt.Data pack (ComfyUI-Manager, Impact Pack).

Common issues & troubleshooting

The preset dropdown loads but generation fails / it can't find models. As the README bluntly puts it, "it is essential for the model's name to be accurate." The helper looks for files by their expected names in your models/ipadapter and models/clip_vision folders. If you renamed a download, it won't find it. Easiest fix: let ComfyUI Manager download the IP-Adapter models so the names match what the preset expects.

You must have cubiq's IPAdapter Plus installed. This node is a wrapper around ComfyUI_IPAdapter_plus. No IPAdapter Plus, no working helper.

FaceID presets error out. FaceID needs the insightface Python package installed - a notorious install headache in its own right - plus the InsightFace model. Set insightface_provider to CUDA (or ROCM on AMD) so it runs on GPU, and make sure the insightface pip package actually installed cleanly.

You picked an SDXL FaceID preset and the likeness is weak. Not this node's fault. The knowledge base is blunt: FaceID on SDXL is experimental and consistently reported as worse than the SD1.5 original. For a real face, do it on SD1.5.