Easy Apply IPAdapter (Regional)

IP-Adapter lets you prompt with an image instead of only words - hand it a reference and it steers style or content toward that picture. The regional version adds the missing dimension: where. Instead of one reference influencing the whole frame, you paint a mask and say "this reference goes here." Two characters with two different looks, a specific style on the left and another on the right - that's the job this node exists for.

It's part of ComfyUI-Easy-Use's Adapter set, and it's the convenience wrapper over the regional IP-Adapter pattern that used to take a small pile of nodes to build.

Why you'd reach for it

Because a single global IP-Adapter smears one reference across everything, and sometimes you want control. The classic case is a two-subject scene where each subject should look like a different reference image. You mask subject A, apply reference A; mask subject B, apply reference B; and because these nodes chain, you build up as many regions as you need. IP-Adapter conditions what things look like while a ControlNet conditions where they go - pairing regional IPAdapter with regional prompts and a ControlNet is one of the stronger composition workflows on SD1.5/SDXL.

How it works

You give it a reference image, a mask for the region it should affect, and per-region positive/negative text. It encodes the reference and limits its influence to the masked area, merging that into the conditioning carried on the pipe. The key to multi-region work is the optional_ipadapter_params input/output pair: it outputs an IPADAPTER_PARAMS bundle that you feed into the next regional node's optional_ipadapter_params input, stacking region after region into one coherent apply.

The inputs that matter

• image - the reference for this region.
• mask (optional input) - where this reference applies. No mask = whole frame.
• image_weight - how strongly the reference pulls, default 1.0. IP-Adapter is sensitive here; back it off if the reference is overriding your prompt.
• prompt_weight - balances the per-region text against the image.
• weight_type - 14 modes including style transfer, composition, strong style transfer. Pick style transfer when you want the look but not the layout.
• start_at / end_at - the sampling window.

Outputs: pipe, ipadapter_params (chain into the next region), positive, negative.

Installing it

Install Easy-Use (ComfyUI Manager → ComfyUI Easy Use, or clone + install requirements). IP-Adapter itself is a separate dependency the pack drives but doesn't bundle:

• cubiq/ComfyUI_IPAdapter_plus custom nodes.
• The IP-Adapter model files and the matching CLIP-vision encoder for your architecture (the model must match your base - SD1.5 vs SDXL).

Where people trip

The most common one is a model/architecture mismatch: an SD1.5 IP-Adapter on an SDXL model (or vice versa) just fails or produces garbage - match them. Second, IP-Adapter softens detail when it runs across all steps, so if your regions look mushy, end them earlier with end_at rather than running full-length. Third, for chaining regions, remember the params handoff - feed each node's ipadapter_params output into the next node's optional_ipadapter_params input, or your later regions won't combine. And note cubiq's pack is in maintenance-only mode and has no Flux support; regional IPAdapter here is an SD1.5/SDXL technique.