☁️BizyAir IPAdapter

What it is

IP-Adapter is what image prompting looked like before Flux-era models started doing it natively. Tencent AI Lab built it as a tiny 22M-parameter adapter - a fraction of the size of a full fine-tune - that reads a reference image and injects that appearance into generation through its own dedicated cross-attention path, while your text prompt keeps working normally alongside it. No LoRA training required; you just hand it a picture. This node is BizyAir's cloud version, wired to run against a BIZYAIR_MODEL on their hardware instead of yours.

Worth knowing where it fits historically: IP-Adapter is squarely an SD 1.5/SDXL-era tool. It's still the default for that generation of models, but newer Flux-class architectures generally route the same "give me the look of this reference" job through different mechanisms (Redux, PuLID, in-context editors like Kontext) rather than IP-Adapter itself. If your loaded model upstream is SD 1.5 or SDXL-family, this is exactly the tool you want; if it's Flux, check whether BizyAir has a more architecture-appropriate node before reaching for this one.

How it works

The adapter encodes your reference image into an embedding and runs it through cross-attention layers dedicated to image conditioning, separate from the frozen text-conditioning path. That separation is the whole trick - the base model doesn't need retraining, and your text prompt isn't fighting the image prompt for the same attention mechanism.

Inputs and outputs that matter

model and ipadapter are both typed sockets from upstream BizyAir loader nodes. image is your reference picture - the one whose style or subject you want carried into generation. weight (default 1, range -1 to 3) is the main strength dial; community convention for style transfer tends to sit around 0.6–0.8, dropping lower if the reference is overpowering your prompt. start_at and end_at (both 0–1) let you limit the adapter's influence to a portion of the sampling process - running it only in early steps is a common way to preserve more fine detail than running it the whole way through. weight_type picks between standard, prompt is more important, and style transfer - self-explanatory names, and worth trying prompt is more important first if the reference image is steamrolling your text description.

The optional attn_mask lets you restrict the adapter's effect to a specific region via a MASK, rather than applying it across the whole image. The single output, m, is a BIZYAIR_MODEL - the original model with the adapter now baked into its conditioning, ready to feed straight into a BizyAir sampler.

Installing it

Search "BizyAir" in ComfyUI Manager and install, or clone directly: cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, then restart. Every node in this pack needs a BizyAir API key configured first - the in-app login prompt handles registration on first run.

Common issues

The single biggest gotcha carries over from IP-Adapter's own history: it was built for SD 1.5 and SDXL, and running it against a Flux or other newer architecture loaded through BizyAir is likely to produce weak or broken results even though the node will happily accept the connection. If your reference image's style is overwhelming everything else, lower weight before you touch anything else - that's the single highest-leverage knob. And if faces or identity specifically aren't holding up well, know that IP-Adapter's face-focused variant (FaceID) is a different beast with its own licensing quirks; this simple version conditions on general appearance, not facial geometry.