chx_IPA_faceID_adv

Same relationship as chx_IPA_adv to chx_IPA_basic: this is chx_IPA_faceID with the preset dropdown replaced by raw file selectors. If you're running a FaceID variant the pack's five presets don't cover, or you need to pin a specific InsightFace model pack rather than let a preset choose for you, this is the node.

What's different from the preset version

Two things move from "picked for you" to "you pick it": the adapter/encoder files, and the InsightFace face-recognition model itself.

• ipadapter_file / clip_vision default to kolor_ipa-faceid-plus.bin and kolors-faceid-clipversion.bin - the FaceID pair built for Kolors. Swap them for whatever FaceID checkpoint and matching CLIP vision encoder your architecture actually needs; like the plain chx_IPA_adv, these dropdowns are populated from whatever's sitting in your model folders, not a curated list.
• insightface here is an explicit choice between buffalo_l and antelopev2 - the two InsightFace model packs. buffalo_l is the default and the only one InsightFace auto-downloads; antelopev2 is more accurate and is what tools like InstantID expect, but it has to be fetched by hand from a manual download link and unzipped into your InsightFace models directory yourself. That manual step, on a link that's broken about as often as it works, is the single most common reason a FaceID-advanced workflow that "should just work" doesn't.

Everything else - the mechanism (an ArcFace identity embedding plus a companion LoRA, patched into the model's attention), and the license situation - is identical to chx_IPA_faceID: FaceID's weights are non-commercial because they depend on InsightFace's non-commercial models, full stop, regardless of which node loads them.

The inputs and outputs that matter

• model, image - base model, face reference.
• ipadapter_file, clip_vision - manual file pickers, as above.
• insightface - buffalo_l or antelopev2.
• weight (default 1, -1 to 3) and weight_faceidv2 (default 1, -1 to 5) - identity strength and the FaceID-v2-specific CLIP-structure weight, same two-dial setup as the preset node. Community range for the companion LoRA weight is 0.5–1.0.
• weight_type, combine_embeds, embeds_scaling - the standard advanced attention/combination controls shared across the whole IPA family.
• start_at / end_at - step range to apply the adapter over.
• Optional attn_mask, image_negative.

Outputs are model, an image output (note: this node's own schema names it iamge, a typo in the pack rather than a naming choice you need to match - since ComfyUI wires by type and position, not by label text, it doesn't actually affect how you connect it), and mask.

Installing it

Through ComfyUI Manager: search ComfyUI-Apt_Preset, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

then install.bat on Windows for the base pack's dependencies, restart. Separately, you need insightface installed and working, plus your chosen FaceID adapter/encoder files and, if you picked antelopev2, that model pack downloaded manually - none of this is auto-fetched by the node itself.

Common issues

antelopev2 doesn't load, or the download link is dead. This is a widely reported problem across the whole InsightFace-dependent tooling space, not an Apt_Preset bug - the official distribution for that pack lives behind a manual download, and links to it break periodically. If it's not working, fall back to buffalo_l (auto-downloaded, no dead-link risk) unless you specifically need antelopev2's extra accuracy.

Import or dependency errors on the whole pack. Confirm this is a base-install issue before troubleshooting this specific node - install.bat (or a pip install of insightface by hand) needs to have actually succeeded first. InsightFace 1.0 dropped the old C++ build-tools requirement, so if you're still being told to install Visual Studio build tools, check you're not on a stale insightface version.

Output identity is close but not exact. Expected ceiling for this family, not a bug: adapter-based identity conditioning is good enough for stylized work and consistently reported as short of true photorealistic likeness. If you need pixel-level accuracy, this is the point where people layer a face-swap pass on top rather than pushing the weight higher.