PixlStash Likeness Search

Semantic search finds pictures by meaning; PixlStash Likeness Search finds pictures that look like a reference. You feed it one or more query images, and it ranks your vault by visual similarity - whole-image CLIP embedding comparison by default, or ArcFace facial-feature comparison when you want "whose face is this" instead of "what looks similar." It's the node behind the pack's own demo of combining concepts: an old man and a young man drinking beer in, and the result is older men drinking beer out.

Two search modes

• picture_likeness (default) - full-image CLIP embedding comparison. Good for style, scene, composition.
• face_search - ArcFace facial features. PixlStash extracts the most prominent face from each query image and matches against faces in the vault. This is your "does this face exist in my library / who is it closest to" tool.

How the pool-and-sample thing works

The clever part is that the node doesn't just hand you the top N. It builds a candidate pool of pool_size top-similar pictures, then returns select_count from it:

• If select_count is smaller than pool_size, it draws a random sample from the pool - so you get variety instead of 10 near-identical frames of the same best match.
• If select_count ≥ pool_size, you get the top pool_size results directly.

Defaults are pool 50, select 10 - a curated-but-varied handful. Raise select_count toward pool_size when you want every strong match; keep them far apart when you want to mine a vein of variations.

threshold (default 0) is the minimum cosine similarity to enter the pool at all. combine decides how multiple query images' scores merge per candidate: mean, max, min, harmonic_mean, or geometric_mean. The README's multi-concept example - combine two references to get "both, roughly" - is mean or max doing that work.

Inputs and outputs

Required: search_mode, image (a batch - every frame is uploaded and combined), combine, pool_size (1–2000), select_count (1–500), threshold. Optional: pixlstash_project / pixlstash_set / pixlstash_character filters.

Outputs: image, mask, pass-through filters, batch_size. Standard shape for this pack - wire image anywhere an IMAGE batch goes.

Install and gotchas

Part of the pack: ComfyUI Manager → ComfyUI-PixlStash, or clone into custom_nodes, restart, set URL/token in Settings > PixlStash. Deps: requests + Pillow. Requires PixlStash v1.4 (development releases for now). No matches → clear error, so if you get one, lower threshold or loosen the filters before blaming the node.

One honest caveat: likeness is rank order, not truth. "Closest to the reference" ≠ "matches the reference." If you need a hard accept/reject verdict rather than a ranked pool, that's what the gates are for.