☁️BizyAir IPAdapterAdvanced

IP-Adapter is what you reach for when you want a generation to look like a reference image - style, composition, or a face - without training a LoRA. It works by extracting a CLIP embedding from your reference and injecting it through its own cross-attention path, separate from the text path, which is why prompts keep working normally alongside it. This node is the "advanced" tier of cubiq's well-known ComfyUI IP-Adapter implementation, running against BizyAir's cloud instead of a local checkpoint - hence the BIZYAIR_MODEL typing rather than plain MODEL.

The inputs that actually matter

model (BIZYAIR_MODEL, from a BizyAir checkpoint loader), ipadapter (from BizyAir IPAdapter Unified Loader upstream), and image (your reference photo) are the core three. weight (-1 to 5, default 1) is the strength dial - worth flagging that the community's real sweet spot for IP-Adapter generally sits lower than this node's default, around 0.3–0.6 for most style-transfer work; push much past 0.7–0.8 and the reference starts overriding your prompt rather than complementing it.

weight_type is where "Advanced" earns its name - 15 options covering linear application plus a family of eased curves (ease in, ease out, ease in-out, reverse in-out) and strength-distribution presets (weak input, strong middle, and so on), plus two purpose-built modes: style transfer and composition, which are cubiq's presets for pulling just the stylistic feel or just the layout from a reference without dragging its exact content along. start_at/end_at (0–1 each) gate which portion of the denoising steps IP-Adapter is active for - useful for letting early steps establish composition freely before the reference kicks in, or the reverse.

combine_embeds matters once you're feeding more than one reference into image as a batch - concat, add, subtract, average, or norm average decide how multiple embeddings get merged into one signal. embeds_scaling (V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty) is a deeper internal knob controlling how the image embedding interacts with attention keys versus values - leave it at default unless you're already deep into IP-Adapter tuning and know what you're chasing.

The optional inputs round it out: image_negative lets you tell the adapter what not to pull from, attn_mask restricts the effect to a region, and clip_vision lets you override the vision encoder feeding it - pair it with BizyAir Load CLIP Vision upstream if the default doesn't match what your IP-Adapter checkpoint expects.

Output

A patched model (BIZYAIR_MODEL) - wire it straight into your sampler the same way you'd use any conditioned model output.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Or:

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

Restart after. No IP-Adapter weights to download - they live cloud-side. A BizyAir account and API key are required before this node runs.

Where people get burned

Starting at weight 1 and wondering why the prompt stopped mattering. That's the node's default, but it's on the high end of what the community actually recommends - dial it back toward 0.3–0.6 first.

Mismatched clip_vision. If results look noisy or subtly off rather than clearly wrong, check whether the vision encoder feeding this node actually matches what the loaded IP-Adapter checkpoint expects.

Reaching for embeds_scaling or weight_type presets without a reason. These are fine-tuning knobs for people already fighting a specific problem - for a first pass, linear and default V only scaling are the sane starting point.