Nunchaku Pulid Loader (Deprecated)

Before anything else: NunchakuPulidLoader is deprecated. Use NunchakuPuLIDLoaderV2 for new work. This node is the original PuLID loader for Nunchaku, kept around because older workflows still call it, and it's worth understanding what it set up.

PuLID is the tuning-free way to put a specific face onto Flux without training a LoRA - ByteDance's identity adapter that copies a face while breaking prompt adherence less than the adapters before it. This loader was the setup step: it prepared a Nunchaku model for identity injection and produced the pipeline object the apply node needed.

How it works

The loader takes your Nunchaku MODEL and returns two things: the model patched for PuLID, and a PULID pipeline object that the (also deprecated) NunchakuPulidApply consumes with your reference image. Compared to the V2, this original path had a leaner interface - it didn't surface things like the EVA-CLIP encoder selection or the InsightFace device choice that the V2 exposes. That's part of why the V2 exists: more control and a cleaner PuLID integration.

Like all PuLID, it leans on InsightFace under the hood to turn a reference face into an identity embedding, which drags in InsightFace's dependencies and its non-commercial licensing.

The inputs and outputs that matter

• model (MODEL) - your Nunchaku model. This is the only input.

Outputs:

• MODEL - patched for PuLID.
• PULID - the pipeline object, which feeds into NunchakuPulidApply.

Note the output type is PULID here, versus the V2's PULID_PIPELINE - the two generations aren't wire-compatible, so don't try to mix an old loader with a new apply node.

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. PuLID's real dependencies apply: the pack's backend wheel, a quantized Flux model, InsightFace (install 1.0+ so you skip the C++ build) with the antelopev2 models, and the PuLID weights.

Common issues & troubleshooting

Migrate to the V2. The single most useful fix: rebuild on NunchakuPuLIDLoaderV2 + NunchakuFluxPuLIDApplyV2. They give you the EVA-CLIP and InsightFace-device controls this node lacks, and they're the maintained path.

InsightFace install pain. The perennial headache. Use InsightFace 1.0+, keep it in ComfyUI's own Python, and fetch antelopev2 by hand into ~/.insightface/models/ - it isn't auto-downloaded.

Mismatched apply node. This loader outputs PULID, which only the deprecated NunchakuPulidApply accepts. If you paired it with the V2 apply node, that's your error - match the generations, or better, move both to V2.

Selling output? InsightFace weights are non-commercial and PuLID inherits that. Plan accordingly.