Load Flux IPAdatpter

Yes, the display name really does say "IPAdatper" - that's XLabs' own typo, not a rendering bug on this page. Don't let it throw you; the node works, the spelling just didn't get a second pass.

What this node actually does is the boring-but-necessary first half of a two-node combo. On its own it produces nothing you can look at - it just loads the model files and hands you a bundled object (ipadapterFlux) that you then plug into Apply Flux IPAdapter or Apply Advanced Flux IPAdapter to actually condition a generation. Think of it as the "open the file" step before "do something with the file."

The point of IP-Adapter as a technique, in case you haven't used one before: instead of typing "a woman with red curly hair in a leather jacket," you hand the model a photo and it borrows the look from that instead. It works by running your reference through a CLIP vision encoder and feeding those image embeddings into the model's attention alongside your text prompt, so the picture and the words steer generation together rather than one replacing the other. Worth knowing going in: this is XLabs' own from-scratch IP-Adapter for Flux, not a port of the classic Tencent IP-Adapter you may have used on SD1.5 or SDXL - Tencent never shipped Flux weights, so every Flux-era "IP-Adapter" (this one, InstantX's, BFL's own Redux) is a separate implementation trained by a different team, and they don't behave identically to what you're used to.

The inputs that matter: ipadatper is which XLabs IP-Adapter checkpoint to load (populated from your controlnet/ipadapter model folder), clip_vision is which CLIP vision encoder to use for reading the reference image, and provider just picks CPU or GPU for that encoder - leave it on GPU unless you're VRAM-starved, since CPU inference here is slow. The single output, ipadapterFlux, is an IP_ADAPTER_FLUX object that only makes sense wired into one of the two Apply nodes downstream.

Installing it: grab the pack through ComfyUI Manager by searching "x-flux-comfyui," or the manual route - cd ComfyUI/custom_nodes && git clone https://github.com/XLabs-AI/x-flux-comfyui then restart ComfyUI. Manager handles the setup step for you; doing it by hand means running python setup.py from inside the cloned folder once. For this specific node you additionally need two downloads: an OpenAI CLIP ViT-L/14 vision model dropped into ComfyUI/models/clip_vision/, and the actual IP-Adapter checkpoint from XLabs-AI/flux-ip-adapter on HuggingFace dropped into ComfyUI/models/xlabs/ipadapters/ (the pack auto-creates that folder on first launch, so it should already be there waiting).

Where people get burned: the authors shipped this thing labeled beta and meant it. Early community reaction on release was mixed to lukewarm - people coming in expecting IPAdapter-Plus-on-SDXL results were disappointed, with reports of the reference influence being noticeably weaker than that, plus a contrast/sharpness shift some users worked around with a color-match pass afterward. None of that is this loader node's fault specifically, it's the underlying adapter weights being an early v1 release rather than a mature one - go in expecting to fight for a good result rather than getting one on the first try. If the empty dropdowns are the actual problem (no options show up for ipadatper or clip_vision), that just means the model files aren't in the right folders yet - this node reads its choices straight off disk. And if generation is oddly slow, double-check provider isn't set to CPU by accident.