☁️BizyAir IPAdapterModelLoader

What it is

IP-Adapter made "image prompting" cheap: instead of training a LoRA to teach a model what something looks like, you hand it a reference photo and a small adapter (about 22 million parameters, a fraction of a fine-tune) injects that image's content through its own attention path, leaving your text prompt free to keep doing its job alongside it. This loader is the first node in that chain - it points BizyAir's cloud at a specific IP-Adapter checkpoint so the rest of your workflow (an apply node, a style-composition node, whatever) has a model to work with.

One detail worth flagging in the schema itself: the sample checkpoint listed is kolors/ip_adapter_plus_general.bin, which tells you BizyAir's IPAdapter support isn't limited to the usual SD 1.5/SDXL lineup - it reaches into Kolors-family checkpoints too, which is a base you won't find covered in most IP-Adapter documentation elsewhere.

The inputs that matter

Just one:

• ipadapter_file - which checkpoint to load. The available options depend on what's hosted on BizyAir's side; expect variants organized by base model family (the Kolors example above is the schema's own sample choice).

Output is a single IPADAPTER handle. Notably this one isn't BIZYAIR_-prefixed like most of this pack's internal types - worth knowing if you're trying to figure out what else in your graph it can plug into.

How to install it

Through ComfyUI Manager: search "BizyAir," install, restart. Or clone it yourself:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

A BizyAir account and API key are required before this - or anything downstream - actually runs; the checkpoint lives on SiliconFlow's infrastructure, not locally, so there's no model file to hunt down and place in a folder yourself.

Common issues

Picking the wrong variant for your base model. IP-Adapter checkpoints are architecture-specific - an SD 1.5 file won't work sensibly against an SDXL pipeline, and neither will work against Kolors. Match the loaded file's family to whatever your generation model actually is before debugging anything downstream.

Weight guessing. This loader doesn't expose a weight control itself - that lives on whatever apply node consumes the IPADAPTER output. As a starting point from general IP-Adapter usage: 0.6-0.8 for style transfer, 0.3-0.5 if you're after face similarity without a dedicated identity adapter. Push past about 0.7-0.8 and the reference image starts overriding your prompt rather than complementing it.

Face-specific work looking soft. Plain IP-Adapter (as opposed to a FaceID variant) reads general visual appearance, not facial geometry - it's genuinely good at style and rough likeness, genuinely weak at nailing an exact face. If precise identity matters more than style, you want an identity-specific tool (InstantID, elsewhere in this same pack) rather than pushing this one's weight higher and higher.

Commercial use. The base CLIP-driven IP-Adapter variants are the commercially clean half of this tool family - it's specifically the FaceID line (face-recognition-based) that inherits InsightFace's non-commercial restriction. If your loaded checkpoint is a plain style/appearance adapter rather than a FaceID one, that restriction doesn't apply to it.