Nunchaku IP-Adapter Loader

Image prompting - hand the model a picture and say "make it look like this" - is what IP-Adapter does, and NunchakuIPAdapterLoader is the setup half of doing it on a Nunchaku 4-bit model. It takes your MODEL, prepares it for IP-Adapter conditioning, and returns both the MODEL and an IPADAPTER_PIPELINE that the apply node uses to actually inject the reference.

IP-Adapter is the cheap alternative to training a LoRA: instead of teaching the model what something looks like, you show it a reference and a small adapter injects that image's embedding through its own attention path, leaving the text prompt working alongside. That decoupling is the whole trick. This loader wires that machinery onto the quantized Nunchaku model, which the stock IP-Adapter nodes don't know how to do.

How it works

The loader's job is preparation: it patches the model so it can accept image-prompt conditioning and packages the adapter setup into an IPADAPTER_PIPELINE. The heavy lifting - encoding your reference image and applying it - happens next, in NunchakuFluxIPAdapterApply. Think of this node as flipping the model into "ready for image prompting" mode and producing the handle the apply step needs.

Worth setting expectations: IP-Adapter conditions on what things look like (style, appearance), where ControlNet conditions on where things go (pose, edges, depth). They're complementary, not competing. If you want structure control, that's a different tool.

The inputs and outputs that matter

This node is deliberately minimal - one input:

• model (MODEL) - your Nunchaku model, from the Flux loader (or downstream in your chain).

Outputs:

• MODEL - the model prepared for IP-Adapter.
• IPADAPTER_PIPELINE - the setup object that NunchakuFluxIPAdapterApply consumes.

You pass both forward. There are no strength or file knobs here on purpose - those live on the apply node, which is where you feed the actual reference image and dial its influence.

How to install it

Part of the ComfyUI-nunchaku pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. Beyond the pack's usual backend wheel and quantized Flux model, IP-Adapter needs its adapter weights and the image encoder they pair with - download the ones Nunchaku's IP-Adapter docs point to and place them in the folders it specifies.

Common issues & troubleshooting

Node loads but does nothing visible. That's expected - this is only the setup half. You won't see any effect until NunchakuFluxIPAdapterApply feeds it a reference image downstream.

Missing adapter/encoder files. If loading fails, you're likely missing the IP-Adapter weights or the image encoder. Grab both from the sources in the pack's IP-Adapter guide.

Don't confuse it with PuLID. For a specific face, PuLID (NunchakuPuLIDLoaderV2) is usually the stronger tool; IP-Adapter is better for general style/appearance transfer from a reference. Pick by the job.

Nothing happens on a non-Nunchaku model. These nodes target Nunchaku's quantized models. On a stock Flux, use the regular IP-Adapter node packs instead.