Li et al. (2023) — ARXIV-2312.04461
Key takeaways
- Fuses an arbitrary number of reference images into one stacked ID
embedding, fused with the embedding of a class word (“man”, “woman”), so
personalization needs no per-subject optimization at inference — unlike
fine-tuning (
LIT-079) or per-subject embedding search (LIT-078) - Because the identity representation is a set, more references give a more complete identity, and the paper claims both fidelity and diversity improve — unusual, since those normally trade, and unexplained
- A capability falls out of the representation: the stack is assembled from one identity during training, but nothing requires that at inference, so feeding images of different people merges identities. Reported as the side effect it is
- An ID-oriented data construction pipeline is presented as part of the method rather than as preliminaries
Standing in the anthology
Read — NOTE-038. Retagged to adaptation-and-tuning. The furthest paper in
this pass from the record’s subject — human-face personalization on a
purpose-built dataset — and nothing is sourced to it.
The durable observation is that identity merging is a second instance of a
pattern LIT-097 also shows: a representation chosen for economy delivering an
unplanned capability. Two independent data points is an argument for preferring
uniform representations over special-cased ones.
Of the takeaways replaced, “style consistency” is not something this paper claims, and none of them mentions that the embedding is a stack of arbitrary size, which is the mechanism.
Audited by #121; unsourced. R1 — prefer a set-valued conditioning representation when the number of references is not fixed — is the transferable one, and the record has no practice about the arity of a conditioning signal. Rated moderate by the reading, single-source, and in a narrow domain; recorded as a candidate rather than filed.