Most identity-preserving generators struggle when asked to place many specified people in one scene: models either fail to assign references to distinct locations or fall back to trivial copy-paste of the reference face. This paper argues that explicit identity–layout grounding — not only embedding matching — is the key to scaling identity-preserving synthesis to larger groups while keeping controllability over placement and appearance.
Key Findings
- Explicit planning raises reliable identity binding: the method predicts a structured identity–layout plan (who goes where) and renders it as a visual condition, so identities are addressed to locations before synthesis.
- Layout-Grounded ID Loss stabilizes supervision: supervising generated faces against annotated face regions avoids noisy embedding-based matching and reduces unstable correspondence among multiple predicted faces.
- Quantitative gains: on an identity-disjoint benchmark, target-context face similarity improves from 0.462 (GPT-Image-2) to 0.499; copy-paste artifacts drop from 0.169 to 0.055; coverage of requested identities reaches 97.3% with a duplicate rate of 2.8%.
- Auxiliary training objective (ID Representation Forcing) and staged tuning (reconstruction, paired tuning, quality tuning) encourage models to predict identities before rendering and to avoid low-level duplication shortcuts.
How it works
- Addressed tokens: each selected identity is injected as an addressed token so the model can attend to identities explicitly.
- Plan-first pipeline: the model predicts a structured identity–layout plan, converts it into a visual conditioning map, and then synthesizes the image conditioned on that plan.
- Loss design: Layout-Grounded ID Loss uses annotated face boxes to assign identity supervision to specific spatial regions; ID Representation Forcing trains identity predictions prior to image decoding; combined with contrastive-style negatives this balances fidelity vs. variation.
Who it's for & trade-offs
Great fit if you need consistent, controllable multi-person synthesis (e.g., group photos, multi-character scenes) and can provide or leverage identity-annotated data. Look elsewhere if you need a zero-shot system that binds many unknown people without any paired or region annotations — the approach depends on annotated face regions, paired references, and training that explicitly breaks copy-paste shortcuts.
Where it fits
Compared with embedding-only or reconstruction-heavy approaches, this work shifts supervision from post-hoc embedding matching to layout-grounded identity supervision, making it particularly relevant for systems that must place dozens of identities precisely and reliably in a single composition.