Load IPAdapter Flux Model

If you've used IP-Adapter on SD 1.5 or SDXL before, the name here is doing you a favor and a disservice at the same time. It's the same idea - hand the model a reference image instead of (or alongside) a text prompt - but it is not the same weights, not the same encoder, and not the same team. This node loads InstantX's FLUX.1-dev-IP-Adapter, a from-scratch adapter built for Flux and shipped through Shakker Labs' ComfyUI pack, distinct from cubiq's SD1.5/SDXL IP-Adapter plus family. Load it here, then feed the result into one of the two Apply nodes from the same pack to actually patch your model.

What it does and why you'd reach for it

Flux never got an official IP-Adapter from Black Forest Labs the way SD 1.5 and SDXL did from Tencent. What it has instead is BFL's own Redux (style/variation), PuLID-Flux (face identity), and this - InstantX's community-built adapter, released November 2024. This node is step one of using it: it loads the adapter checkpoint plus the vision encoder that reads your reference image, and hands both downstream as a single bundled object.

Worth knowing going in, straight from the model's own card: it's explicitly not built for fine-grained style transfer or tight character consistency. It's a broader-strokes tool - general appearance and composition guidance - and the community mostly treats it that way. Reddit chatter on Flux IP-Adapter is a mixed bag: some people get results they call "stunning," others say flatly that IP-Adapter peaked in the SD 1.5 era and for faces specifically you're better off with PuLID-Flux or ACE++. If your goal is locking in a specific face, skip ahead to one of those. If you want a reference image to loosely steer style, composition, or mood alongside your prompt, this is the tool, and it's cheap to try.

The inputs that matter

Three fields, all dropdowns:

• ipadapter - pick the adapter checkpoint file. It only shows up here after you've downloaded it (see install below); an empty list means the file isn't in the right folder yet.
• clip_vision - the image encoder. This adapter uses SigLIP (google/siglip-so400m-patch14-384), not the OpenCLIP encoder classic IP-Adapter uses - that's the core architectural swap that makes this a different tool under a familiar name.
• provider - cuda, cpu, or mps. This is which device runs the vision encoder. Leave it on cuda if you've got an Nvidia GPU (which, running on ComfyICU, you do); mps is for Apple Silicon, cpu is the fallback if nothing else works.

The output

A single ipadapterFlux output of type IP_ADAPTER_FLUX_INSTANTX. That's not a MODEL - it's a bundle wire that plugs directly into ApplyIPAdapterFlux or ApplyIPAdapterFluxAdvanced, which is where the actual model patching (and the weight/percent knobs) happen. This loader by itself doesn't touch your model at all.

Installing it

Via ComfyUI Manager: search ComfyUI-IPAdapter-Flux, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Shakker-Labs/ComfyUI-IPAdapter-Flux
cd ComfyUI-IPAdapter-Flux
pip install -r requirements.txt

Then the model files - these don't come with the node pack:

1. Download the InstantX FLUX.1-dev-IP-Adapter weights into ComfyUI/models/ipadapter-flux.
2. Get google/siglip-so400m-patch14-384 into ComfyUI/models/clip_vision, or just let it pull from the HuggingFace cache on first run.

Restart ComfyUI so both dropdowns populate. The pack's own README also flags a minimum ComfyUI version - if you're on an install that's genuinely old, update ComfyUI core first.

Where people get burned

The empty ipadapter dropdown is the number one support question in disguise - it's not a bug, it just means the weight file isn't in models/ipadapter-flux yet, or it's in the wrong subfolder. Second: this pack installs its own requirements.txt into the same Python environment as every other custom node you have, and version conflicts (transformers, in particular) are the most common way a fresh custom-node install goes sideways in ComfyUI generally - if the loader throws an import error on startup, check the console log before assuming the node itself is broken. And the licensing note that's easy to miss: the adapter weights ship under the FLUX.1-dev Non-Commercial License, separate from the pack's own Apache 2.0 code - fine for personal work, worth checking if you're building something commercial.