chx_IPA_adv

chx_IPA_basic gives you a preset dropdown (STANDARD, PLUS, PLUS FACE, and so on) and picks the right adapter and CLIP vision files behind the scenes. chx_IPA_adv skips that abstraction - you point it directly at an ipadapter_file and a clip_vision file yourself. It's the same IP-Adapter mechanism, same attention-patching trick, just with the model-selection training wheels off.

Why you'd reach for the "adv" version

The preset system on chx_IPA_basic covers the common cases, but if you're running a model the pack's presets don't map cleanly to - the node's own defaults point at kolor_ip_adapter_plus_general.bin and kolors_clip_version.bin, which are the IP-Adapter files built for Kolors (Kwai's SDXL-based model) - you need to be able to swap in a different file pair by hand. This node is that escape hatch: any IP-Adapter checkpoint and CLIP vision encoder you've downloaded into the right folders shows up in the dropdowns, not just the pack's curated presets.

Worth restating from the wider IP-Adapter picture: this whole family covers SD 1.5, SDXL, and Kolors. It does not exist for Flux - no version of IP-Adapter, from Tencent AI Lab or otherwise, loads on that architecture, so this node won't help you there regardless of which file you point it at.

The inputs and outputs that matter

• model and image - base model and reference picture, same as the basic version.
• ipadapter_file / clip_vision - the raw file pickers. Whatever's populating your models/ipadapter and models/clip_vision folders shows up here; match the encoder to the adapter file (a Kolors adapter wants the Kolors CLIP vision encoder, an SDXL Plus adapter wants ViT-bigG-14, and so on - mismatching them is the single most common reason this style of node produces garbage).
• weight (default 1, -1 to 5) - same role as in the basic node: how strongly the reference pushes. 0.6–0.8 for style transfer, 0.3–0.5 if you're chasing face similarity without the dedicated FaceID line.
• weight_type, combine_embeds, embeds_scaling - the same set of advanced curve/combination controls as the basic node; leave at default unless you know what you're tuning for.
• start_at / end_at - restrict the adapter's influence to a step range. Earlier-only application keeps more detail; running the full range softens the image.
• Optional image_negative (steer away from a second reference) and attn_mask (restrict the effect spatially).

Outputs: model (patched), image, 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 dependencies, restart. This node needs actual IP-Adapter checkpoint files and a matching CLIP vision encoder placed in models/ipadapter and models/clip_vision before the dropdowns will show anything useful - the pack doesn't bundle them. The Apache 2.0 CLIP-based weights live on Hugging Face under h94/IP-Adapter; Kolors-specific files (matching this node's defaults) come from the Kolors project directly.

Common issues

Dropdowns are empty. The ipadapter_file and clip_vision schema for this node shows zero built-in choices by default - meaning it's populated entirely from what you have on disk. If nothing shows up, you haven't put any files in models/ipadapter or models/clip_vision yet, or ComfyUI hasn't rescanned since you added them (restart, or use your Manager's model-folder refresh).

Output is broken or garbled. Almost always an adapter/encoder mismatch. IP-Adapter files are trained against a specific CLIP vision model (OpenCLIP ViT-H-14 for SD1.5-era files, ViT-bigG-14 for SDXL, a Kolors-specific encoder for Kolors files) - pairing the wrong two together is the classic failure here, and unlike the preset-driven basic node, nothing stops you from doing it by accident.

You're not sure which files you need. If you don't have a specific reason to hand-pick files, start with chx_IPA_basic instead - its preset system exists precisely to avoid this class of mistake.