All Prompt Attention SEGS

Its sibling Concept Attention SEGS makes you type which concepts you care about (girl | pink hair | cat). This node removes the typing. Point it at the image that came out of a sampler and it reads the positive prompt's mapped tokens, then returns separate, overlapping soft SEGS for every concept the model actually attended to while generating. If you've ever stared at a generation and wondered which words were doing the work - or wanted regions for every subject without enumerating them - this is the discovery tool.

How it differs from Concept Attention SEGS

Both are SimpleSyrup's attention-capture nodes: they read cross-attention recorded during sampling, trace the graph back to the sampler that made the image, and turn attention into labeled SEGS. The differences are what make each worth having:

• No concepts input. The concept list comes from the positive conditioning's mapped tokens automatically. There is genuinely no concepts field on this node - look at the schema and you won't find one.
• evidence_mode defaults to raw attention. Concept Attention SEGS defaults to "concept isolation," which applies weighting to derive a cleaner stable region. Here you see the captured model probabilities more directly, which is better for understanding what the model was doing and worse if you want tidy silhouettes. Switch it to concept isolation when a region is headed into real downstream work.
• The output list is as big as your prompt. Every mapped concept gets its own SEGS, overlapping where concepts overlapped in the model's attention.

The three outputs, and the controls that tame them

• image passes through unchanged.
• segs is the list - one labeled entry per concept instance.
• mask is the soft union of every retained region.

Since "every concept" can be a crowd on a busy prompt, the controls that matter here are the pruning ones. keep_only (default 1) keeps the best N instances per concept, ranked by keep_by (default largest size, or attention confidence if you prefer). minimum_region_size discards attention components under 512 pixels by default - raise it when you're drowning in specks. combine_segs merges all retained instances of each concept into a single SEG. sampler_stage picks which sampler in the chain you're reading, and capture_start/capture_end trim which slice of denoising counts as evidence. The rest (minimum_strength, minimum_consensus, geometry_recall, split_sensitivity, instance_recall, matte_solidity, edge_feather, capture_profile) are the shared family stack, sensibly defaulted.

Why you'd actually reach for it

Three situations. First, you want to see what your prompt is doing - which tokens drew real attention and where, as an inspectable SEGS map. Second, you're building an automatic regional workflow and don't want to hand-maintain a concept list that changes every prompt; this node makes the regions a pure function of whatever prompt you wrote. Third, the SEGS output is Impact-compatible, so the regions feed detailers and the pack's regional samplers exactly like Concept Attention SEGS - but now they scale to any prompt without edits.

Watch out for

Same provenance rule as the whole family: the image must trace back to the sampler in the same workflow, untouched in between, or there's no captured attention to read. And because evidence starts as raw attention, expect overlap - the same pixels legitimately belong to "girl" and "pink hair" - which is deliberate, not a bug. If you're feeding these into something that wants a tight single region, switch evidence_mode to concept isolation and lean on minimum_strength and minimum_region_size. For a quick look, though, defaults plus the segs output are exactly right.

Part of SimpleSyrup: install once via Manager (search SimpleSyrup) or git clone into custom_nodes plus requirements.txt. No model downloads - the model was already doing the attention while it generated.