Feature Dashboard

Once Feature Atlas has built you a JSON catalog of what every SAE feature does, Feature Dashboard is how you actually read it. It's a browser over that atlas: search features by keyword, look at category distributions, explore co-occurrence clusters, rank by selectivity or activation strength - and every mode spits out the feature indices you're hunting, ready to paste into Feature Gate or Feature Probe. It turns "20,000 anonymous numbers" into "searchable library you can shop from."

How it works

You hand it an atlas JSON (it also accepts a plain Feature Dictionary JSON) and pick a mode. Each mode is a different lens on the same data:

• overview - category distribution + health summary. Start here.
• search - find features by keyword in their labels. This is the money mode: try lighting, color, texture, dark and get every feature whose top-activating prompts mention it.
• clusters - co-occurrence cluster visualization. See which features hang out together (likely redundant or related).
• top_selective - features that fire 5–30% of the time. The tooltip calls these "most interesting" - selective enough to mean something, common enough to matter.
• top_activation - highest mean activation.
• top_rare - rarest features.
• top_norm - strongest decoder columns.

Every mode outputs FEATURES, the paste-ready index list. So a typical research loop is: Dashboard → search "golden" → copy three indices → Feature Probe to verify what they do → Feature Gate to apply. Search proposes, probe confirms, gate applies.

The inputs that matter

• atlas_path - the atlas JSON from Feature Atlas (or a Feature Dictionary JSON).
• mode - which view to render.
• search_keyword - only used in search mode. The tooltip's suggestions (lighting, color, texture, dark) are worth trying first.

Outputs: VISUALIZATION (IMAGE) and FEATURES (STRING).

Installing

Standard for the pack - ComfyUI Manager search "Concept Steer", or clone https://github.com/Nynxz/ComfyUI-ConceptSteer into custom_nodes/ and restart. This one is cheap to run: it just reads a JSON you already built, so no encoder, no VRAM blowout, no downloads. The only real requirement is an atlas to browse.

Common gotchas

• Search is only as good as your labels. If you skipped Feature Dictionary / Feature Atlas labeling, features have no text to match and search finds nothing. Build the dictionary first.
• "Selective" and "strong" are different things. top_selective and top_activation answer different questions - a feature can fire rarely but hard, or often but weakly. Know which you're picking.
• Indices are meaningless outside their SAE. The dashboard is browsing one specific atlas; those indices only gate/probe correctly against the same SAE (same layer, expansion, seed).