Nunchaku FLUX IP-Adapter Apply

This is where the reference image actually goes in. NunchakuIPAdapterLoader gets IP-Adapter ready; NunchakuFluxIPAdapterApply takes that setup plus your reference picture and injects its look into a Nunchaku 4-bit Flux, handing back a MODEL that generates in that style. It's the business end of image prompting on Nunchaku.

The pitch for IP-Adapter over training: no LoRA, no dataset, no wait. You show the model one image, and a lightweight adapter carries its appearance into your generation through a separate attention path, so your text prompt still works next to it. Fast, cheap, and reversible.

How it works

The node encodes your image with IP-Adapter's image encoder into an embedding, then applies that embedding to the model through the adapter's cross-attention, weighted by weight. The base model stays frozen - you're adding an image-conditioning signal alongside the text one, not overwriting the model. Out comes a patched MODEL for your sampler.

The key dial is weight, and IP-Adapter has a well-known behavior worth internalizing: crank it too high and the reference overrides your prompt. On classic IP-Adapter, people run style transfer around 0.6–0.8 and back off when the text stops mattering. Treat that as your mental model here.

The inputs and outputs that matter

• model (MODEL) - the prepared model from NunchakuIPAdapterLoader.
• ipadapter_pipeline (IPADAPTER_PIPELINE) - the other output of that loader.
• image (IMAGE) - your reference. This is the picture whose look you're borrowing.
• weight (default 1.0, range 0–5, step 0.05) - how strongly the reference influences the output. Start around 0.6–1.0. Lower keeps your prompt in charge; higher makes the output hug the reference and can flatten prompt detail.

Output is a MODEL - into your KSampler.

How to install it

Ships with ComfyUI-nunchaku. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

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

then restart. It's downstream of NunchakuIPAdapterLoader, so the same requirements carry over: the backend wheel, a quantized Flux model, and the IP-Adapter weights + image encoder the loader needs.

Common issues & troubleshooting

The reference barely shows up. Raise weight. If it's still weak, confirm the loader actually produced a valid pipeline and that you fed it a decent, on-topic reference image.

Output ignores my prompt. Classic IP-Adapter over-weighting - pull weight down toward 0.5–0.7 so the text regains control.

Softer, less detailed output. Image-prompt conditioning across the whole generation tends to soften detail. That's inherent to the approach; a lower weight helps, and so does letting the model do a detail pass.

Wanted a face, got a vibe. IP-Adapter transfers general appearance, not precise identity. For a specific person, use PuLID (NunchakuFluxPuLIDApplyV2) instead - it's built for faces.