Load IPAdapter Flux Model (Advanced)

Straight up: the widgets on this node are the same three you'll find on the plain IPAdapterFluxLoader - ipadapter, clip_vision, provider. If you're expecting a pile of extra sliders because it says "Advanced," you won't find them here. What you will find is a matching "Advanced" apply node downstream (ApplyIPAdapterFluxAdvanced) that has real extra controls - a weight ramp instead of a flat value, plus a step count. This loader is best understood as the loader half of that advanced pairing, not as a more configurable loader on its own.

That said, don't over-think the pairing: both loaders in this pack output the exact same type, IP_ADAPTER_FLUX_INSTANTX, and both Apply nodes accept that same type on their ipadapter_flux input. So this loader will happily feed the plain ApplyIPAdapterFlux too, and the plain loader will happily feed ApplyIPAdapterFluxAdvanced. They're interchangeable at the wiring level - pick whichever Loader/Apply combo matches how much control you want over the weight, not because one loader is locked to one apply node.

What it loads

Same as its sibling: InstantX's FLUX.1-dev-IP-Adapter, the Shakker Labs-distributed community adapter for Flux built on a SigLIP image encoder rather than the OpenCLIP encoder classic IP-Adapter (cubiq's SD1.5/SDXL pack) uses. It's a from-scratch Flux build, not a port - nothing from the original Tencent IP-Adapter loads on Flux at all. Context worth carrying over from the rest of the pack: the model card itself says it isn't tuned for fine-grained style transfer or tight character consistency, and community sentiment on Flux IP-Adapter generally is split - genuinely good results for some people, "IP-Adapter peaked at SD 1.5" from others who'll point you at PuLID-Flux or ACE++ for anything face-critical.

The inputs

• ipadapter - the adapter checkpoint dropdown, populated once you've downloaded the weights into the right folder.
• clip_vision - the SigLIP encoder (google/siglip-so400m-patch14-384).
• provider - cuda, cpu, or mps, the device running the vision encoder. cuda unless you have a specific reason not to.

The output

ipadapterFlux, type IP_ADAPTER_FLUX_INSTANTX - wire it into either Apply node. This loader doesn't touch your diffusion model; it just bundles the adapter and its encoder for the Apply step to consume.

Installing it

Same install as the whole pack - there's no separate install path for the Advanced loader specifically. 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 grab the two model files the README calls out: the InstantX FLUX.1-dev-IP-Adapter weights into ComfyUI/models/ipadapter-flux, and google/siglip-so400m-patch14-384 into ComfyUI/models/clip_vision (or let it cache from HuggingFace on first use). Restart ComfyUI so the dropdowns populate.

Where people get burned

The ipadapter dropdown shows up empty until the checkpoint file is actually sitting in models/ipadapter-flux - that's not a broken node, that's a missing download. If you searched for "advanced" settings expecting more knobs than the basic loader and came up empty, that's expected too - the extra control in this pack lives on the Apply side, not the Loader side, so head to ApplyIPAdapterFluxAdvanced for the weight scheduling. And the same general custom-node landmine applies here as anywhere: this pack's requirements.txt installs into your shared ComfyUI Python environment, and dependency version clashes (transformers especially) are the most common reason a fresh node install throws errors on startup - check the console log rather than assuming the node itself is at fault.