Regional IPAdapter Encoded Mask (Inspire)

This is the "I already did the encoding" version of regional IP-Adapter. Its siblings take a reference image and a mask and handle everything; this one takes a plain MASK and pre-computed EMBEDS - the CLIP Vision embedding of your reference - and applies that to the masked region. If you've already encoded your reference (say with cubiq's IPAdapter Encoder, or you're blending several references into one embed), you hand the embeds straight in here instead of an image.

Why bother? Two real reasons. First, control: encoding separately lets you combine multiple reference images into a single embed before regional application, or reuse one expensive encode across several regions. Second, it's the path for embeds you've manipulated - averaged, weighted, or built from a batch - that never existed as a single image.

Same era caveat as the rest of the family: this is an SD1.5/SDXL workflow. Per the knowledge base, IP-Adapter has no Flux-generation successor in adapter form, so regional image prompting lives on those older architectures.

How it works

You give it a mask marking the region and the embeds to apply there. It binds those embeds as an IP-Adapter conditioning scoped to the mask, and emits a REGIONAL_IPADAPTER handle. Chain one per region, combine them, apply the combined result in the sampler. The difference from the image-based nodes is purely the front end: instead of "image → encode → apply," you've already got the encode and you're jumping straight to apply.

The inputs and outputs that matter

• mask - a MASK (not a color map) defining where these embeds apply. You bring your own, from a mask editor, a ColorMapToMasks node, or anywhere else.
• embeds - the pre-encoded EMBEDS for your reference. This is the whole reason to use this node.
• weight (default 0.7) - strength of the reference in-region. KB guidance: ~0.6–0.8 for style, lower if the prompt gets drowned out.
• start_at / end_at (0 to 1) - the denoising window the adapter runs over. End early to keep detail; running IP-Adapter across all steps softens the output.

There's also weight_type, unfold_batch, and an optional neg_embeds if you're supplying a negative reference embedding.

The single output is REGIONAL_IPADAPTER - combine it with the other regions' handles and feed the sampler.

How to install it

ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

then restart. It's a Dr.Lt.Data pack (ComfyUI-Manager, Impact Pack).

Common issues & troubleshooting

weight_type says "IPAdapterPlus is not installed." That literal string in the dropdown means the dependency is missing. These nodes ride on cubiq's ComfyUI_IPAdapter_plus - install it via Manager and restart, and the real options appear. Without it the node is inert.

Where do the embeds come from? From an IP-Adapter encoder node (cubiq's pack has one). If you don't already have an EMBEDS source in your graph and just want to point at an image, use Regional IPAdapter Mask or Regional IPAdapter By Color Mask instead - those take an image and encode it for you.

Embed / model mismatch. Embeds are tied to a specific CLIP Vision encoder, which is tied to your base model. Embeds encoded for an SD1.5 setup won't behave on SDXL. Keep the whole chain - encoder, adapter, base - on one architecture.

Regions bleed. Standard regional-conditioning behavior. Shorten the masking window (end_at) so hard masks only fix composition and the base blends the rest - the KB's consistent advice for avoiding the collaged, seamy look.