Runware IP-Adapter

IP-Adapter is the classic "show, don't tell" conditioning: feed it a reference image and it steers generation toward that image's look - style, composition, or both - through decoupled cross-attention, no training required. Locally it's a legacy stack that never followed the ecosystem past SDXL, which is exactly why this node is interesting. It's a builder in Runware's pack, so it runs IP-Adapter on Runware's cloud against the pack's SDXL-era architecture nodes - the same trick, minus the model files, the cubiq dependency tree, and the "does this load on my checkpoint" gamble.

Think of builders as the pack's config nodes: they produce a typed value (here, RUNWARE_IPADAPTERS) that you wire into a model node's ipAdapters socket. It doesn't generate anything by itself - it's the seasoning, not the meal.

The inputs that matter

• model (required): which IP-Adapter variant, from four choices (runware:56@1 through @4). Runware's catalog lists them; if in doubt, the first is the usual general-purpose pick.
• guideImages (required, IMAGE): the reference look. One or more images.
• weight (default 1, range −4 to 4): influence strength. 0 means no influence; negative values steer away from the reference - genuinely useful for "anything but this." For style transfer, the classic move is low positive weight, around 0.3–0.5.
• weightType (13 choices): how influence evolves over the generation. The interesting ones are style_transfer, composition, strong_style_transfer, and style_and_composition - pick by which half of the image you care about. Everything from ease_in to strong_middle shapes the curve.
• combineMethod: how multiple guideImages combine - concat, add, subtract, average, norm_average. subtract is the niche trick for "reference minus another reference."
• embedScaling: which embedding components get used (only_v, kv, ...) - leave it alone unless you know why you're here.
• ipAdapters: the chain input. Builders are stackable, so wire one IP-Adapter's output into the next's ipAdapters socket to apply several sequentially.

The weightComposition toggle (gated) reveals a value for controlling composition influence specifically - the "keep the layout, ignore the colors" fine-tune.

How it works

The builder collects your settings into an IP-Adapter descriptor, and when the model node runs, that descriptor rides along in the API request - stripped to only the fields the target model accepts. The adapter runs cloud-side against the checkpoint. Because it's a builder, the output is just the typed value; wire it into any model node that exposes an ipAdapters socket.

Install and gotchas

Pack install is one-time: ComfyUI Manager → search Runware, or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes + pip install -r requirements.txt, restart, and set a Runware API key.

Two things to internalize. First, this only helps on models that accept ipAdapters - that's the SD 1.5/SDXL-era architecture nodes like RunwareArch_sd_1_5, not the Flux-generation nodes, because IP-Adapter simply didn't get rebuilt for DiT. Second, a disconnected or empty builder silently produces None rather than an error, so if nothing's happening, check your guideImages actually has an image. And remember the stack is paid - every adapter in a chain is part of one generation's cost, so stack deliberately, not decoratively.