Nunchaku Pulid Apply (Deprecated)

Straight up: NunchakuPulidApply is deprecated - the current node is NunchakuFluxPuLIDApplyV2. This is the original apply step in Nunchaku's PuLID setup, documented here because older graphs and guides still use it.

Its job is the payoff of the whole PuLID chain: take the pipeline from NunchakuPulidLoader plus a reference photo, and inject that person's identity into a Nunchaku 4-bit Flux, returning a MODEL that generates them. PuLID's whole appeal is getting a specific face without training a LoRA, and keeping your prompt working while it does.

How it works

You feed it the PULID pipeline, an image (the reference face), the model, and a strength. Internally the reference runs through InsightFace for an identity embedding, and PuLID conditions the diffusion model on it during sampling. Out comes a patched MODEL. The V2 version added more control over when and how identity is applied (start/end scheduling, an attention mask); this original keeps it to a single strength dial.

The realistic expectation is the same across versions: PuLID nails the face structure and stops a few percent short of exact likeness. The community recipe is PuLID first, then a face-swap finishing pass to close the gap.

The inputs and outputs that matter

• pulid (PULID) - the pipeline from NunchakuPulidLoader.
• image (IMAGE) - the reference face; use a clean, front-facing shot.
• model (MODEL) - your Nunchaku model.
• ip_weight (default 1.0, range 0–2, step 0.01) - identity strength. 1.0 is the starting point; up for stronger likeness, down if the face overrides your prompt.

Output is a MODEL for 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. As the tail of a PuLID chain it needs everything that chain needs: the backend wheel, a quantized Flux model, InsightFace (1.0+) with antelopev2, and the PuLID weights.

Common issues & troubleshooting

Use the V2. NunchakuFluxPuLIDApplyV2 gives you start/end scheduling and a mask input, so you can end identity injection early for a more natural finish. Rebuild on it.

"No face detected." InsightFace couldn't find a face in your reference - use a larger, clearer, front-facing image.

Likeness too weak or too strong. That's ip_weight. Note its range here caps at 2, narrower than the V2's weight - another reason the V2 gives you more room.

Wrong node pairing. This apply node takes the PULID type from the deprecated loader, not the V2's PULID_PIPELINE. Keep both nodes on the same generation, or move the whole chain to V2.