Attention Capture Model (Internal)

The display name is doing real work here: Attention Capture Model (Internal). This is a node you will see in pack listings and registry searches, that has basically zero community footprint, and that will confuse you for exactly one reason - you can't find a good reason to wire it yourself. That's because there isn't one. It's the behind-the-scenes machinery for SimpleSyrup's attention-region family (Concept Attention SEGS, All Prompt Attention SEGS, Attention Region Mask, Attention Masked Conditioning), and the pack wires it for you.

What it actually does

Those region nodes work by reading the cross-attention that happened while the image was being sampled - which spatial pixels attended to which prompt tokens, and when. To capture that, something has to sit on the model path of the sampler and record it. That something is this node. It clones the upstream MODEL, attaches one observation-only attention observer to the clone, and hands the patched model back so sampling proceeds unchanged except for the recording.

Its inputs are the contract for that job:

• model - the upstream model, cloned for observation-only capture.
• positive - the positive conditioning whose prompt tokens get mapped to image regions.
• plan_json - a string containing the capture plan for the target sampler: which denoising stages to observe, which concepts to isolate, how dense the observation should be.
• clip (optional) - makes prompt-to-token mapping exact when the graph can see the CLIP model.

The single output is a MODEL you'd route into a sampler. Sounds like something you could just do manually, right? And that's the trap.

Why you don't add it by hand

plan_json is not a field you fill in. It's a serialized plan generated by the pack's own graph planner, and the format is internal. SimpleSyrup registers a queue-time hook in ComfyUI that rewrites your workflow graph before validation: when you add an attention-region node, the hook walks your graph backwards from that node, finds the sampler that produced the image, and injects an AttentionCaptureModel into that sampler's model path with a correctly-built plan. The node is marked dev-only - it's deliberately kept out of the normal "add node" browse experience because manual placement would just produce broken plans.

So the practical answer to "how do I use this node" is: you don't, and you shouldn't. Add a Concept Attention SEGS or Attention Region Mask, connect an image or latent that traces back to a sampler, and this node appears in the executed graph as the pack's scaffolding. If you ever find a stale one sitting orphaned in a workflow, delete it - the region nodes rebuild it on the next queue.

Where it fits in the wider pack

It's the clearest example of how far SimpleSyrup's regional machinery goes compared to older packs: instead of detecting regions after the fact with a YOLO or SAM model, it captures evidence during generation and turns that into masks and SEGS afterward. Nothing here needs a detector download, which is why this internal node exists at all. Understanding that the region nodes lean on live attention capture also explains the one real constraint they all share: the image or latent you connect has to trace back to the sampler in the same workflow, untouched in between. Break that provenance chain and there's nothing for the observer to have recorded.

Installation is just installing SimpleSyrup once - see any of the pack's other node pages for the exact steps (Manager search "SimpleSyrup", or clone into custom_nodes and install requirements.txt; you'll want a current ComfyUI since the pack uses the V3 extension API). No models, no extra downloads for this one.